Finding rule groups to classify high dimensional gene expression datasets

Abstract

Microarray data provides quantitative information about the transcription profile of cells. To analyze microarray datasets, methodology of machine learning has increasingly attracted bioinformatics researchers. Some approaches of machine learning are widely used to classify and mine biological datasets. However, many gene expression datasets are extremely high dimensionality, traditional machine learning methods cannot be applied effectively and efficiently. This paper proposes a robust algorithm to find out rule groups to classify gene expression datasets. Unlike the most classification algorithms, which select dimensions (genes) heuristically to form rules groups to identify classes such as cancerous and normal tissues, our algorithm guarantees finding out best-k dimensions (genes) to form rule groups for the classification of expression datasets. Our experiments show that the rule groups obtained by our algorithm have higher accuracy than that of other classification approaches.

Introduction

Mining gene expression datasets has generated interest among many bioinformatics researchers (Alon et al., 1999, Chen et al., 2007, Knapp and Chen, 2007, Mann et al., 2007, Mramor et al., 2007). One of the important trends in bioinformatics is identification of genes or groups of gene to differentiate diseased tissues from normal tissues. Classification of tissues into cancerous and normal tissues using the identified genes is one of the key problems being faced in bioinformatics. Golub (Golub et al., 1999) firstly showed the better diagnostic performance of gene expression signatures in acute leukemia classification compared to other currently used diagnostic method. Many other studies (Bhattacharjee et al., 2001, Khan et al., 2001, Shipp et al., 2002) have been undertaken in almost all cancer types. Recently, a wide range of statistical and machine learning methods for microarray data analysis developed (Allison et al., 2006, Nahar et al., 2007, Asyali et al., 2006, Pham et al., 2006). Since the particularity of microarray data, i.e. high number of genes and small number of samples, it becomes a hard task to find accurate patterns to classify microarray data. This paper attempts to find rule groups for several different cancerous datasets. The results are encouraging in terms of accuracy and effectiveness.

Gene expression data is usually represented as a matrix; each element in the matrix represents an appearance level of a particular gene under a particular condition. We assume that a gene expression matrix has n rows and m columns. The rows represent samples that are divided into different classes such as cancerous tissue and normal tissue. The columns represent genes whose number is usually more than several thousands. The number of rows is much lower than that of columns as the sample used ranges from ten to several hundreds. To cope with this kind of extremely high dimensional data, traditional machine learning techniques such as decision tree and support virtual machine, cannot classify effectively as they use heuristics to select significant dimensions (genes); many discriminative dimensions can be left out. In this paper, we propose a classification method that generates rule groups to categorize samples. A rule is a conjunction of several dimensions (genes); each gene is constrained into one interval. For example, (gene1 > 120.5) ∧ (gene2 ≤ 20.3) is one such rule. If a sample satisfies the conjunction of a rule, it will be covered by the rule. The above rule covers samples whose expression values of gene1 are larger than 120.5 and expression values of gene2 are smaller than or equal to 20.3. In contrast to traditional machine learning algorithms that use heuristics our method guarantees finding out best-k genes which are most discriminative to classify samples in different classes, to form rule groups. The value of parameter k is set to around 5. It is based on the fact that each rule should not be too long from the principle of Occam's razor (Mitchell, 1997); otherwise, the problem of overfitting will arise (Quinlan, 1986).