Novel protein-function prediction using a directed hybrid graph

Abstract

Proteins carry out various important activities in an organism. Accurately annotating their functions can boost the advance of life-science research and application. High-throughput techniques generate such a large volume of proteomic and genomic data that it is beyond the capability of low-throughput wet-lab based techniques. Thus, computational model-based large-scale protein-function prediction is one of the key tasks in the post-genomic era. Current machine-learning based methods often focus on predicting the functions of completely unlabeled proteins. These methods ignore the incomplete labels of the labeled proteins, and hence have low accuracy. In this paper, we design a directed Hybrid Graph (dHG) based on the gene ontology hierarchy and the protein-protein interaction network. Next, we use the dHG to predict novel functions by performing a random walk with restart on it. The proposed dHG can predict not only new functions for partially labeled proteins, but also new functions for completely unlabeled proteins. Experimental results on proteins of yeast and humans show that dHG, across various evaluation metrics, achieves better results than other related methods, and costs less time than these methods.