MultiG-Rank: Multiple graph regularized protein ranking

Background Protein ranking is a fundamental task in structural biology. Most protein ranking methods rely on the pairwise comparison of proteins while neglecting the global manifold structure of the protein database. Recently, graph regularized ranking is proposed by exploiting the global structure of the graph defined by these pairwise similarities. However, the existing graph regularized ranking methods are very sensitive to the choices of graph model and parameters, which remains a tough problem. Results To solve this problem, we have developed Multiple Graph regularized Ranking algorithm - MultiG-Rank. Instead of using a single graph to regularize the ranking scores, we approximate the intrinsic manifold of protein distribution by combining multiple initial graphs for the regularization. The graph weights are learned with ranking scores jointly and automatically, by minimizing an object function alternately in an iterative algorithm. Experimental results on the ASTRAL SCOP protein database demonstrate that MultiG-Rank achieves the better ranking performance comparing to both other single graph regularized ranking methods and pairwise similarity base ranking methods. Conclusion The problem of graph model and parameter selection in graph regularized protein ranking can be effectively solved by combining multiple graphs. This aspect of generalization introduces a new frontier in applying multiple graph to solving protein ranking applications.