Incorporating heterogeneous biological data sources in clustering gene expression data

In this paper, a similarity measure between genes with protein-protein interactions is proposed. The chip-chip data are converted into the same form of gene expression data with pearson correlation as its similarity measure. On the basis of the similarity measures of proteinprotein interaction data and chip-chip data, the combined dissimilarity measure is defined. The combined distance measure is introduced into K-means method, which can be considered as an improved K-means method. The improved K-means method and other three clustering methods are evaluated by a real dataset. Performance of these methods is assessed by a prediction accuracy analysis through known gene annotations. Our results show that the improved K-means method outperforms other clustering methods. The performance of the improved K-means method is also tested by varying the tuning coefficients of the combined dissimilarity measure. The results show that it is very helpful and meaningful to incorporate heterogeneous data sources in clustering gene expression data, and those coefficients for the genome-wide or completed data sources should be given larger values when constructing the combined dissimilarity measure.