Discriminatory Mining of Gene Expression Microarray Data

Recent advances in machine learning and pattern recognition methods provide new analytical tools to explore high dimensional gene expression microarray data. Our data mining software, VISual Data Analyzer for cluster discovery (VISDA), reveals many distinguishing patterns among gene expression profiles, which are responsible for the cell's phenotypes. The model-supported exploration of high-dimensional data space is achieved through two complementary schemes: dimensionality reduction by discriminatory data projection and cluster decomposition by soft data clustering. Reducing dimensionality generates the visualization of the complete data set at the top level. This data set is then partitioned into subclusters that can consequently be visualized at lower levels and if necessary partitioned again. In this paper, three different algorithms are evaluated in their abilities to reduce dimensionality and to visualize data sets: Principal Component Analysis (PCA), Discriminatory Component Analysis (DCA), and Projection Pursuit Method (PPM). The partitioning into subclusters uses the Expectation-Maximization (EM) algorithm and the hierarchical normal mixture model that is selected by the user and verified "optimally" by the Minimum Description Length (MDL) criterion. These approaches produce different visualizations that are compared against known phenotypes from the microarray experiments. Overall, these algorithms and user-selected models explore the high dimensional data where standard analyses may not be sufficient.