Research On Extracting Drug-drug Interactions From Texts Based On External Knowledge

Drug-drug interactions(DDIs)occur when patients are given multi-kinds of drugs,the effects of the drugs may be enhanced or weakened by other drugs,or serious adverse drug reactions(ADRs)may occur.Therefore,the study of the relation between drugs has been widely concerned by the biomedical community.The study of drug-drug interaction can provide guidance for patients to take drugs and provide reference for drug development.In recent years,the model of drug relation extraction based on deep learning has replaced the feature-based machine learning method and become the mainstream model of biomedical relationship extraction.Compared with traditional machine learning methods such as support vector machine(SVM),deep neural network does not need complex artificial features,and has better generalization performance.Integrating external knowledge into deep neural network is a hot topic in recent years.External knowledge can help deep neural network better understand the semantic information of related entities,and then improve the performance of neural network model.Knowledge graph is a kind of external knowledge,which contains rich information of entities and relationships between entities.The pre-trained language model represented by BERT is another hotspot in recent years.Neural network model can achieve better results in most natural language processing tasks by substituting the input of traditional word vector with the pretrained language model.In this research,a method using drug description documents as external knowledge is proposed to enhance the neural network understanding of drug names.An output-modified bidirectional transformer(BioBERT)and a bidirectional gated recurrent unit layer(BiGRU)are used to obtain the vector representation of sentences.Vectors of drug description documents encoded by Doc2 Vec are used as drug description information,which is an external knowledge to our model.Then three different kinds of entity-aware attentions are constructed to get the sentence representations with entity information weighted,including attentions using the drug description information.The outputs of attention layers are concatenated and fed into a multilayer perception layer.Finally,the result are got by a softmax classifier.The F-score is used to evaluate our model,which is also adopted by most previous DDIs extraction models.The proposed model is evaluated on the DDIExtraction 2013 corpus,which is the benchmark corpus of this domain,and achieves the F-score of 80.9%.Then,this research studies how to integrate multiple drug-related knowledge databases,and extracts drug-related information from them to help distinguish DDI relation,and proposes a method of integrating drug information from knowledge graph into neural network model for classification.In this method,multiple biomedical knowledge bases are integrated into a knowledge graph,and then a neural network is used to train drug vectors according to drug relations.The trained drug vectors and the final MLP layer of DDI classification network are spliced,and then the classification results are obtained by softmax classifier.The classification network of DDI is similar to the model proposed above.The model achieves 81.2% F-score on DDIExtraction 2013 corpus.In conclusion,this research proposes a relation extraction model based on BioBERT and two kinds of external knowledge for the problem of drug relationship extraction in biomedical relationship extraction.Experiments on DDIExtraction 2013 data set show the effectiveness of the proposed models.The knowledge graph in this research contains a wealth of drug-related information,which can be used in other biomedical tasks.