Visualisation of Reduced-Dimension Microarray Data Using Gaussian Mixture Models

Dimensionality reduction, clustering and visualisation methods proposed in recent years have afforded new possibilities for the analysis of gene expression data. However, efficient, novel techniques for processing and representing microarray data are still required. We propose the use of the discrete cosine and sine transformations for dimensionality reduction of microarray data. These techniques have found powerful applications in the signal processing domain. Gaussian mixture models (GMMs) are then used for clustering and visualisation of the reduced-dimension data. Results on human fibroblast microarray data reveal that the discrete sine and cosine transforms can greatly reduce the dimensionality of gene expression data while preserving good clustering results. GMMs are shown to produce improved clustering results according to an intra-class cluster tightness criterion, in addition to a new twodimensional representation whose axes afford the possibility of physical interpretations.