Building Text-mining Framework for Gene-Phenotype Relation Extraction using Deep Leaning

Abstract

The scientific literature is a rich resource for information retrieval on the biological knowledge. Nevertheless, the unstructured textual data in the research articles makes it difficult to access the information with computer-aided systems. Text-mining is one of the solution that can transform unstructured information in the text into database content, and most of the approaches are based on the machine learning models. Since these approaches require high-dimensional features, the performance of the model is heavily dependent on the selection of features. However, it is usually difficult and labor-intensive to choose good features, because feature extraction requires prior knowledge and ingenuity of human experts. Here, we suggest a novel framework to extract biological relations from the texts by using hierarchical text features that enhance the effectiveness of relation extraction model.

The proposed framework is composed of two parts, node and edge detection, using deep belief networks. Each part is based on the hierarchical text features learned by Gaussian-Bernoulli restricted Boltzmann machine (GBRBM). In this work, we performed gene-cancer relation extraction task as a pilot study. The classification model was trained based on both GE09 corpus from BioNLP'09 Shared Task and CoMAGC corpus. The results show that our model achieved better performance than other handcrafted feature-based approaches. The evaluation results suggest that deep belief networks offers the optimized and generalized hierarchical text features for the large-scale text mining.