Prediction Of Protein Function Based On Weighted Random Walk

Abstract:Protein function prediction is a fundamental topic in the post-genomic era, in which a great deal of achievements has been acquired so far. One major branch of computational methods for function annotation is to utilize protein-protein interaction network data. However, the way of assigning protein functions according to protein-protein interaction network data cannot produce good results. It is still a big challenge for researchers designing effective model to make good use of various biological resources. By integrating other sources of biological data, with proper models constructed, the biological information can be effectively adopted to improve the accuracy for protein function prediction.A novel method is proposed to predict protein function based on weighted bi-random walk. In consideration of the fact that there exist a lot of false positives in currently available protein interaction datasets, we use edge clustering coefficient weighting the interactions. We consider the structure of GO terms in the DAG to calculate the GO term similarity. Once the similarities of all GO term pairs are calculated, the functional interrelationship network can be constructed. We adopted three variations of Bi-Random Walk to walk different steps and different sequences in the two networks. We analysis the known protein-GO term to investigate at which level neighbors of proteins tend to have functional associations and at which level neighbors of GO Terms usually co-annotate some common proteins. The experiment results show the comparison with some known classical algorithms, a better predicting performance of our algorithm achieved.We propose a method combine domain information to predict protein function based on weighted thr-random walk. Protein domain is the unit of protein function and structure. Two proteins are likely to share similar functions if they have similar domain compositions. Domain co-occurrence network is constructed in that way, which consists of all the protein domain types of yeast proteins in PPI network as nodes. An edge is introduced to connect two domain types if they co-exist in one protein. In this paper, we adopted a method named ThrRW, which take several steps of random walking on three different biological networks, weighted protein-protein interaction network, domain co-occurrence network and functional interrelationship network respectively so as to get functional information from the neighbors in corresponding networks. The result show that our method achieves better prediction performance not only that than methods that only use PPI network data, but also than methods that consider both PPI data and GO term similarities, which verify the effectiveness of our method on integrating multiple biological networks. Additionally, our method provides a good way to predict functions for protein domains.