The advantage of functional prediction based on clustering of yeast genes and its correlation with non-sequence based classifications

Sequence similarity is probably the most widely used tool to infer functional linkage between proteins. The fully sequenced, much researched, genome of Saccharomyces cerevisiae gives us on opportunity to compare and statistically quantify computational methods based on sequence similarity, which aim to detect such linkage. In addition, the amount of data regarding Saccharomyces Cerevisiae genes and proteins, which is not directly based on sequence is rapidly increasing. Consequently, it allows investigation of the connections and correlation between classification based on these types of data and that based solely on sequence similarity. In this work we start with a simple clustering algorithm to cluster genes based on the BLAST E-score of their similarity. We analyze how well one can infer function from these clusters and for how many of the genes that are currently unknown one can suggest a prediction. Given these parameters, we show that even a simple algorithm achieves better results than simply considering the BLAST output of matching genes. In the second part of the paper, we show that there is a highly significant correlation (p-value, 10^-4 for the vast majority of the experiments) between the aforementioned clusters and other types of classifications. Namely, we show that a pair of genes being clustered together is correlated with these genes having similar expression patterns in DNA array experiments and with the encoded proteins being involved in protein-protein interactions. Although this correlation is highly significant, it is, of course, not strong enough to be, by itself, a tool for predicting co-regulation of genes or interaction of proteins. We discuss possible explanations for this correlation. Furthermore, the statistical evaluation of these results should be considered when developing tools that are aimed at making such predictions.