Ensemble Dependence Model Based Cancer Classification using Gene Microarray Data

DNA microarray technologies make it possible to simultaneously monitor thousands of genes expression levels. A topic of great interest is to study the different expression profiles between microarray samples from cancer patients and normal subjects, by classifying them at gene expression level. Currently, various clustering methods have been proposed in the literature to classify cancer and normal samples based on microarray data, and they are dominantly data-driven approaches. In this paper, we propose an alternative approach, a model-driven approach. We propose an ensemble dependence model, aiming at exploring the group dependence relationship of gene clusters. Under the framework of hypothesis-testing, we employ genes’ dependence relationship as a feature to model and classify cancer and normal samples. The proposed classification scheme is applied to five cancer data sets, and it is noted that the proposed method yields very promising performance. We further analyze the eigen domain of the proposed method, and discovered different patterns between cancer and normal samples. I. BACKGROUND AND PROPOSED SCHEME Current methods for the classification of microarray gene expression data can be mainly divided into two categories. One is based on clustering, which can be used to distinguish cancer and normal samples, and subtypes of cancers. Example schemes include Hierarchical clustering, Local Maximum clustering, Self-Organizing Map, and K-means clustering. These clustering methods are mainly data-driven approaches. Usually, they do not require much prior assumption, i.e., the underlying model. However, determining the number of clusters is a challenging problem itself, and there lacks of widelyaccepted measures to evaluate the clustering performance. The other category is mainly based on machine-learning approach. Motivated by the success of machine learning algorithms in image and speech processing, many researches have been reported to apply them to microarray data analysis. For example, support vector machine and neural network analysis. Machine learning methods generally yield better results than that of the traditional clustering methods. In this paper, we propose an ensemble dependence model based classification approach, as illustrated in Fig 1(a). It includes four main components, feature selection, gene clustering, ensemble dependence model and hypothesis testing. Due to the limited size of current data, it is not feasible to examine the regulation relationship between all genes. Also, the microarray gene expression data is noisy. However, if genes are clustered in a right way, the noise level in the resulting cluster expression will be reduced, and we will be able to reveal the ensemble dynamics of gene clusters. (a) (b) Fig. 1. (a) classification procedure; (b) ensemble dependence model Since not all genes’ expression profiles are informative in understanding the difference between cancer and normal samples, feature selection is needed to exclude irrelevant genes. As mentioned above, gene clustering is performed to group together genes with similar expression. To average out experiment noise and enhance genes’ common expression within each cluster, average gene expression profile is used to represent each cluster. Without any prior knowledge, we assume that, each cluster is to some extent dependent on all the other clusters, as shown in Fig 1(b). Linear dependence relationship is studied, in Equation (1), where aij represents an inter-cluster dependence relationship. The so-called selfregulation is assumed to be zero, i.e. aii = 0, i = 1, 2, 3, 4. Because cluster average is used to represent each cluster, intracluster dependence relationship is averaged out.