Learning Structural Genetic Information via Graph Neural Embedding

Learning continuous vector representations of genes has been proved to be conducive for many bioinformatics tasks as it can incorporate information of various sources including gene interactions and gene-disease interactions. However, most of the existing approaches, following a paradigm stemmed from the natural language processing community, treat the embedding context in a flat fashion such as a sequence, and tend to overlook the fact that proteins are more likely to function together. In this study, we propose an unsupervised gene embedding algorithm which utilizes graph convolutional network to learn structural information of genes from their neighborhoods in genetic interaction networks. We also propose a neighborhood sampling strategy to generate training samples. Our approach does not assume conditional independence of the node neighborhood and focuses on learning structural information. We compare our method against state-of-the-art baselines and experimental results demonstrate the effectiveness of our approach.