Predicting Gene Function in Yeast through Adaptive Weighting of Multi-Source Information

The value of combining informations obtained from different methods, for gene function predictions, has been illustrated by several studies [1, 2, 3, 4]. We propose a new scoring framework, called Adaptive Score (AdS), for predicting the function of a few unclassified Yeast genes. We mainly focus on phenotypic profiles [5], microarray gene expression (All Yeast data [6]), KEGG pathway database [7], protein sequence similarity through transitive homologues, and protein-protein interactions from BioGRID [8] as data-sources. We use the Pearson correlation for similarity extraction from phenotypic profile and gene expression data. All the protein sequences, except Yeast proteins, corresponding to each KEGG pathway (121 pathways in the second level) are downloaded from PIR to extract profile similarity between two Yeast proteins. Profile vector for each protein in Yeast is computed by comparing its sequence across 121 pathway databases, using BLAST. The method is similar to phylogenetic profile [9] construction. To find the similarity between two genes using KEGG profiles, we used the ratio of dot product value and OR value between two profiles. To detect similarity between two proteins sequences through transitive homologues, 37,66,477 protein sequences are downloaded from UniProt and compared with target proteins by using BLAST [10], the metric of ProClust [11], and the method described in [12]. For protein-protein interaction study, manually curated catalogues of known interactions are downloaded from BioGRID [8] and binary interactions are used as the common unit of analysis.