Deep Convolution Neural Networks for Drug-Drug Interaction Extraction

Abstract

When a patient takes two or more drugs within a certain time, the efficacy of one drug may be influenced by the other. This phenomenon is called drug-drug interactions (DDIs). DDIs are important and helpful information for both medical staff and patients to make sure that the drugs co-administrated at the same time have a positive effect on therapy of patients. Many approaches have been applied into drug-drug interaction extraction tasks such as support vector machine (SVM), recurrent neural network (RNN) and long short-term memory (LSTM) in particular. However, the structures of these models are relatively shallow for DDI extraction research compared with the deep neural networks employed in the field of computer vision. However much better results can be obtained with a deep problem-specific architecture which develops hierarchical representations. Hierarchical and deep neural networks may improve DDI extraction. To address this problem, we present a hierarchical and deep neural network to enrich the feature extraction process to enhance the performance of DDI extraction. In this article, we present a deep convolutional neural network (DCNN) based on DDI extraction method. In this method, we firstly apply embedding mechanism to get the semantic and syntactic of the original biomedical literature. Then a novel architecture using small convolutions is proposed, which takes raw biomedical literature as input and operates directly at the word level to get the embedding-based convolutional features. Finally, these features are fed to softmax classifier to extract DDIs from biomedical literature. Our experimental results on the DDIExtraction 2013 corpus show that the performance of network increases as the network gets deeper and hits its peak at depth 16, which obtains a better result (an F1 score $\mathrm {o}\mathrm {f}0.845$) than other state-of-the-art methods.