Construction of phenotypic similarity network of rare diseases and candidate disease gene prediction using RWRHN

Abstract

Recent studies such as text-mining against huge biomedical databases, devising ontology similarity measure, and predicting candidate disease genes using various methods has been applied on the rare disease field. In this study, we aim to devise and compare models of prioritizing candidate rare disease genes. Disease phenotype, which is defined as an every possible observation from patient to diagnose primarily, plays an important role in inferring novel disease related genes recently. Therefore we compute pairwise phenotypic similarity of rare diseases using 3 similarity measures which are Wang, Resnik, and Cosine. The rare disease ontology was extracted from OMIM and Orphanet databases. Next we construct rare disease similarity network where nodes are rare diseases and edges are connected when the similarity value of two disease nodes are higher than cutoff. To group phenotypically similar diseases, we use two strategy which is collecting neighboring disease of query disease and clustering on the rare disease similarity network. Phenotypically similar diseases of query disease is further used as an input to predict potential disease genes. We construct heterogeneous network which consists of physical protein-protein interaction network (PPI) from BioGrid, phenotypic rare disease similarity network, and bipartite disease gene association network obtained from HPO database. In the gene prediction step, we iteratively simulate the random walker on the heterogeneous network using RWRHN algorithm, and the resulting probability of potential candidate rare disease genes to be related to the query diseases are prioritized. We evaluate 6 rare disease gene prediction model which is from a combination of 3 ontology similarity measure and 2 disease grouping methods by employing 10-fold cross validation and by calculating ROAUC. Finally we analyze factors that influence the performance of 6 models.