Pattern recognition approaches in DNA microarray analysis

Abstract

The release of the human genome working draft marked the biomedical discipline opening a new era in the fields of biology and medicine with the use of bioinformatics. In combination with the advent of microarray technology, scientists can now derive a vast amount of valuable information but the need still remains to understand and exploit it. DNA microarray technology allows researchers to study the behavior of thousands of genes in a single experiment, exploring and monitoring their expression in various diseases with the aim of understanding or discovering the biological mechanisms involved. Studying simultaneously this massive gene expression information is a difficult and very demanding task in many ways, making the use of computational and pattern recognition approaches a necessity. Among all thousands of genes studied, many might be irrelevant or redundant to a specific class discrimination problem. This vast amount of data which might contain noise along with redundant information needs to be processed in such a way so that the real valuable and useful knowledge is finally distilled. This “distillate” of marker genes then could be used by an expert to search, discover and understand the hidden biological mechanisms involved in the development of cancer. One may argue that this problem is a typical feature selection paradigm which could be faced efficiently through a number of typical pattern recognition and machine learning approaches. However the problem referred to as ‘curse of dimensionality’ is intensified, which may block the effectiveness of an approach. Usually, in such an application domain each sample (patient) is described though a set of genes, yielding a huge dimensional space (order of thousands) covered by a few (order of tenths) patients. Such a sparse covered space may befool a method to a random or unrealistic solution. The problem is mainly tackled by two types of approaches the filter and the wrapper approach, each one facing the problem from a quiet different perspective: the static one of the filter approach, and the dynamic one through the wrapper approach. Hybridization or integration of these two approaches is an interesting concept which is addressed in the thesis. In assessing the validity of the derived results we, as many others use various stringent statistical criteria. We take the statistical framework one step further, by studying several issues related to the stability and generalization of algorithmic performance. In addition, we expand our evaluation in assessing the biological significance of results achieved through the use of publicly available and widely accepted biological knowledge. This effort contributes in the establishment of an evaluation framework where the various examined methodologies could be tested in a subjective and fair manner leading to both statistical significant and biologically valid results. Integration of biological knowledge is an open problem in the field of marker gene selection. Based on the fact that different research teams propose different solutions, with minimal or no overlap at all among them, we propose an evolutionary process of assembling the biological knowledge contributed by different efforts towards a more unified and global approach.