Clustering Gene-Expression Data: A Hybrid Approach that Iterates Between k-Means and Evolutionary Search

Abstract

Summary. Clustering genes based on their expression profiles is usually the first step in geneexpression data analysis. Among the many algorithms that can be applied to gene clustering, the k-means algorithm is one of the most popular techniques. This is mainly due to its ease of comprehension, implementation, and interpretation of the results. However, k-means suffers from some problems, such as the need to define a priori the number of clusters (k )a nd the possibility of getting trapped into local optimal solutions. Evolutionary algorithms for clustering, by contrast, are known for being capable of performing broad searches over the space of possible solutions and can be used to automatically estimate the number of clusters. This work elaborates on an evolutionary algorithm specially designed to solve clustering problems and shows how it can be used to optimize the k-means algorithm. The performance of the resultant hybrid approach is illustrated by means of experiments in several bioinformatics datasets with multiple measurements, which are expected to yield more accurate and more stable clusters. Two different measures (Euclidean and Pearson) are employed for computing (dis)similarities between genes. A review of the use of evolutionary algorithms for gene-expression data processing is also included.