Localized Co-expressed Pattern Mining On High Dimensional Gene Expression Data

With the development in DNA microarray technology, expression values of thousands of genes can be simultaneously measured efficiently in biological process. Co-expressed gene patterns are essential in revealing gene functions, gene regulations, subtypes of cells, and cellular processes of gene regulatory networks. However, many co-expressed patterns are similar in a group of genes only under specific experimental conditions. In this thesis, we mainly focus on two kinds of localized co-expressed gene patterns: co-attribute pattern and co-tendency pattern.For co-attribute pattern, most of current algorithms are for 2D and 3D datasets. The existing high dimensional frequent closed pattern (FCP) mining algorithm is not very efficiency, especially on dense data. We proposed a high dimensional FCP mining algorithm HDminer. It is based on space partition and inherently has better performance than algorithms based on enumeration tree, especially on dense datasets. We conducted experiments on real gene expression data and synthetic data to show its efficiency and scalability.As to co-tendency pattern, we proposed the first high dimensional localized clustering algorithm, HHLC. It uses hierarchical scheme and facilitates a progressive refinement of results. We conducted an experiment on Arabidopsis gene expression data. We studied the parameters of HHLC and the result indicates our algorithm can find clusters with significant biology interests.