Gene Function Prediction via Discriminative Graph Embedding

Gene function has been a subject of interest but it is far from fully understood. It is known that some genes have certain functions but it is not clear whether those are all the functions they have. It is a recent trend to use different means to predict gene functions; one of them is to use computational methods on large data sets. Different types of information are used in computational methods, each coming with different hypotheses and applicable to different part of genomes. We regard this problem as a prediction problem on networks. The idea is to use gene network information with a hypothesis that a link in the network means a highly probable similarity in gene functions. This is formulated as a node classification problem on graph with smooth label assignment with respect to the graph structure. We propose a method to embed the graph into a Euclidean space considering label information. The embedding is supposed to trade off the graph structure preservation to have a good discriminative ability. Then, label inference can be done via many machine learning techniques, in which we formulate a large margin spectral transform, known for its robustness. We apply the method to predict gene functions on gene network [1] and protein-protein interactions network [2]. Our method gives higher prediction ability than traditional ways of embedding graphs.