De-Noising, Clustering, Classification, and Representation of Microarray Data for Disease Diagnostics

Abstract

A microarray works by exploiting the ability of a given mRNA molecule to bind specifically to the DNA template from which it originated under specific high stringency conditions. After this, the amount of mRNA bound to each DNA site on the array is determined, which represents the expression level of each gene. Qualification of the mRNA (probe) bound to each DNA spot (target) can help us to determine which genes are active or responsible for the current state of the cell. The probe target hybridization is usually detected and quantified using dyes/flurophore/chemiluminescence labels. The microarray data gives a single snapshot of the gene activity profile of a cell at any given time. Microarray data helps to elucidate the various genes involved in the disease and may also be used for diagnosis /prognosis. In spite of its huge potential, microarray data interpretation and use is limited by its error prone nature, the sheer size of the data and the subjectivity of the analysis. Initially, we describe the use of several techniques to develop a pre-processing methodology for denoising microarray data using signal process techniques. The noise free data thus obtained is more suitable for classification of the data as well as for mining useful information from the data. Discrete Fourier Transform (DFT) and Autocorrelation were explored for denoising the data. We also used microarray data to develop the use of microarray data as diagnostic tool in cancer using One Dimensional Fourier Transform followed by simple Euclidean Distance Calculations and Two Dimensional MUltiple SIgnal Classification (MUSIC). To improve the accuracy of the diagnostic tool, Volterra series were used to model the nonlinear behavior of the data. Thus, our efforts at denoising, representation, and classification of microarray data with signal processing techniques show that appreciable results could be attained even with the most basic techniques. To develop a method to search for a gene signature, we used a combination of PCA and density based clustering for inferring the gene signature of Parkinson’s disease. Using this technique in conjunction with gene ontology data, it was possible to obtain a signature comprising of 21 genes, which were then validated by their involvement in known Parkinson’s disease pathways. The methodology described can be further developed to yield future biomarkers for early Parkinson’s disease diagnosis, as well as for drug development.