Negentropy Based Agglomerative Clustering for Microarray Data

Scientists have been successful in cataloguing genes through genome sequencing projects, and they can now generate vast quantities of gene expression data using microarrays. However, due to the sheer size of the data sets involved, and to complexity of the problems to be tackled, the biological comunity has so far had less success in understanding how genes and proteins are connected and how they operate within networks. Such challenges call for a novel approach to data mining and understanding heavily relying on artificial intelligence tools. In this work we propose a hierarchy of two unsupervised clustering algorithms to accomplish clustering of noisy data. The first algorithm is based on a competitive Neural Network or on a Probabilistic Principal Surfaces approach and the second one on an agglomerative clustering based on both Fisher and Negentropy information. In this way a hierarchical clustering algorithm is obtained. For this algorithm different definitions of Negentropy information could be used. We note that the only a priori information that we need is a dissimilarity threshold. From the several tests accomplished on complex data we noted that the approach has good performance. We also note that the dissimilarity threshold is able to understand the data set behavior. The method has been applied to microarray data set of cell cycle in yeast Saccharomyces Cerevisiae. We note that these are noisy gene expression data with missing data points. In this case the reliability of the clusters provided by our method is confirmed by the fact that clusters selected for the analysis contain genes with very low p-values.