Coherent Gene Expression Pattern Finding Using Clustering Approaches

Analysis of gene expression data is an important research field in DNA microarray research. Data mining techniques have proven to be useful in understanding gene function, gene regulation, cellular processes and subtypes of cells. Most data mining algorithms developed for gene expression data deal with the problem of clustering. The purpose of this thesis is to study different clustering approaches for gene expression data. Our first contribution is a dissimilarity measure (DBK) which retains the regulation information and is robust to outliers. We have developed a graph-based clustering algorithm (GCA) for gene expression data. Its main idea is that, inter-cluster genes have more repulsion among them than intra-cluster genes. In particular, at any given moment, genes are clustered based on a repulsion factor which is based on the genes that are yet to be assigned a cluster. This consideration leads to an objective function that is used to find the cluster parameter that optimizes this objective function. Comparison of GCA with competitive algorithms over different real world data sets shows the superiority of our approach. We have also developed a nearest neighbor based clustering algorithm which incorporates frequent itemset mining (FINN). The output of the frequent itemset mining phase is fed as input to the nearest neighbor clustering for detection of clusters. The process is iterated over multiple passes. After each pass, the dataset is pruned by not considering the genes that have already been assigned clusters. Experimental evaluation shows the method is capable in finding finer clustering of the dataset. This thesis also includes a density based clustering algorithm (DGC) which uses the regulation information and the order preserving property of gene expression profiles to cluster genes into high density regions separated by sparse density regions. The proposed algorithm has been validated on several real-life data sets and found to perform well in comparison to similar algorithms. This thesis also incorporates an incremental version of the DGC algorithm (incDGC). Experimental results on six real world gene expression datasets demonstrate that incDGC can cluster the data in an efficient manner while at the same time obtain the same result as when DGC is applied to the whole updated database. All clustering algorithms have been validated using various statistical measures to show their effectiveness