The Research On Gene Expression Profiles Data For Tumor Classification

DNA Microarray is able to detect the expression levels of thousandsof genes simultaneously. Gene expression data collected from DNAmicroarray are characterized by a large amount of genes but with only asmall amount of samples. Based on gene expression profiles, it is of greatimportance to classify the data into different classes at the molecular level.However, the fact that the number of variables is much bigger than that ofthe associated observations makes many classical classification methodsinapplicable. When facing these problem, we focused on studying newmethods to analyze the tumor gene expression data. The main works canbe summarized as follows:Firstly, a novel two dimensional principal componentanalysis(2DPCA) and two dimensional linear discriminantanalysis(2DLDA) based method for tumor classification using geneexpression data was proposed. To begin with, The original DNAmicroarray gene expression data were modeled by Relief-F. Then, according to the selected gene set each1D sample is converted into2Dmatrix sample, and a set of features are extracted by adopting2DPCA anditerative2DLDA. Last, support vector machine (SVM) and K nearestneighbor (KNN) are used to classify the modeling data. This method cansufficiently utilize2D structure information in the process of extractingfeature.Secondly, with the sparse representation for classification of tumorsusing gene expression profiles, namely the meta-sample based robustsparse coding classification (MRSCC), was presented for tumorclassification. we first extracted the meta-samples of every class usingsingular value decomposition (SVD), which may capture alternativestructures inherent to the tumor data and provide biological insight. Thenwe applied robust sparse coding (RSC) to classify the tumor samples usingthe extracted meta-sample, the proposed method is evaluated incomparison with some representative methods, including the SRC andmeta-sample based sparse representation classification.