Clustering gene expression data using a diffraction-inspired framework

The recent developments in microarray technology has allowed for the simultaneous measurement of gene expression levels. The large amount of captured data challenges conventional statistical tools for analysing and finding inherent correlations between genes and samples. The unsupervised clustering approach is often used, resulting in the development of a wide variety of algorithms. Typical clustering algorithms require selecting certain parameters to operate, for instance the number of expected clusters, as well as defining a similarity measure to quantify the distance between data points. The diffraction-based clustering algorithm however is designed to overcome this necessity for user-defined parameters, as it is able to automatically search the data for any underlying structure.

The diffraction-based clustering algorithm presented in this paper is tested using five well-known expression datasets pertaining to cancerous tissue samples. The clustering results are then compared to those results obtained from conventional algorithms such as the k-means, fuzzy c-means, self-organising map, hierarchical clustering algorithm, Gaussian mixture model and density-based spatial clustering of applications with noise (DBSCAN). The performance of each algorithm is measured using an average external criterion and an average validity index.

The diffraction-based clustering algorithm is shown to be independent of the number of clusters as the algorithm searches the feature space and requires no form of parameter selection. The results show that the diffraction-based clustering algorithm performs significantly better on the real biological datasets compared to the other existing algorithms.

The results of the diffraction-based clustering algorithm presented in this paper suggest that the method can provide researchers with a new tool for successfully analysing microarray data.