Computing distributional semantics of genes using graph

Abstract

In order to find out whether a gene causes a certain cancer to progress or regress, an analysis of researches conducted on the relation between the state change of the cancer and the expression change of the gene is called for. However, it is difficult to read articles related to the gene over the network of gene interactions manually when numerous articles are published everyday. To address this problem, we can use text mining tools to find out the relationships between cancers and genes reported in articles, but currently available tools process each sentence differently. It means that they do not take global context such as biological pathways and deductive reasoning into consideration as they cannot detect inter-sentence relations. In particular, since biological pathways often contain many cases where a lot of genes, proteins, and other chemicals interact with one another to perform a task, not considering global context becomes a significant problem. This research proposes a method for computing embedding vectors of vertices on a graph based on Distributional Hypothesis, and uses it to construct the system that infers relationships between genes and cancers by embedding entities like genes from a network of entities on a vector space. After performing experiments using the proposed method, and by considering chains of interactions on a pathway by computing embeddings of entities on a graph, the performance is found to have increased compared to the current system even though the used graph is inaccurate. Additionally, it is found that the decision of whether local constraints are preserved or not when graph embedding is done on a pathway extracted from the literature affects the overall performance.