Predicting gene function by combining expression and interaction data

In this study we combined the spurious protein interaction data from the Database of Interacting Proteins with the recently published gene expression data of S. cerevisiae grown with limited nutrient limitations under different physical/chemical conditions (Tai et al.) in order to predict protein interactions and protein functions with more confidence. Because proteins often have multiple functional annotations, we propose to employ a continuous metric (e.g. the cosine angle) for measuring functional similarity. We show that it is possible to extract multiple functional associations of a gene, but only by applying a strict Pearson correlation threshold on the gene expression data. Using this strategy, we were able to predict the function of six formally unclassified proteins. Additionally, we revealed six small networks of interacting proteins. These networks strongly match with existing biological knowledge. Furthermore, transcription factors could be assigned to four of these interaction networks by employing a recently published transcription database (Harbison et al.).