Effect of feature extraction and feature selection on expression data from epithelial ovarian cancer

Classifying the gene expression levels of normal and cancerous cells and identifying the genes most contributing to this distinction propose an alternative means of diagnosis. We have investigated the effect of feature extraction and feature selection on clustering of the expression data on two different data sets for ovarian cancer. One data set consisted of 2176 transcripts from 30 samples, nine from normal ovarian epithelial cells and 21 from cancerous ones. The other data set had 7129 transcripts coming from 27 tumor and four normal ovarian tissues. Hierarchical clustering algorithms employing complete-link, average-link and Ward´s method were implemented for comparative evaluation. Principal component analysis was applied for feature extraction and resulted in 100% segregation. Feature selection was performed to identify the most distinguishing genes using CART® software. Selected features were able to cluster the data with 100% success. The results suggest that adoption of feature extraction and selection enhances the quality of clustering of gene expression data for ovarian cancer. Identification of distinguishing genes is a more complex problem that requires incorporating pathway knowledge with statistical and machine learning methods.