Clustering Gene Expression Time Series Data

Efficiently and effectively finding the genes with similar behaviors from microarray data is an important task in bioinformatics community. Co-expression genes have the same behavior or are controlled by the same regulatory mechanisms. Clustering analysis is a very popular technique to group the co-expressed genes into the same cluster. One of the key issues for clustering gene expression time series data is to define the similarity between two time series. Distance measurements and correlation coefficients are commonly used similarity definitions. Two time series might be very distant, but they might be similar if a few items are dropped off from one of the two time series. In this paper, we consider this new aspect of time series similarity, denoted “shift effect,” which indicates temporal gap between two time series. For partition based clustering methods, users have to specify the target number of clusters. This is usually done by means of try-and-error to pick up a number from a large range. In order to solve this problem, we apply sequential pattern mining technique by treating time series as sequences. The number of frequent patterns is the number of target clusters. All the time series supporting a sequential pattern are the initial members of a cluster. Then, each time series is iteratively re-assigned to a suitable cluster.