Research On Association Of Biomedical Entities Based On Deep Learning

Despite the rapid development of science and technology,traditional drug discovery is still a time-consuming,labor-intensive,high-risk and high-cost process.Therefore,a new drug development technology is urgently needed to shorten the drug development cycle and reduce risks and costs.Research on the association between biomedical entities(such as genes,diseases,drugs,etc.)is the basis of drug repositioning technology.Research is based on various calculation methods and combined with omics data in various biomedical fields to discover potentially unknown drugs-disease,disease-gene relationship.Due to the superior performance of neural network and recommendation model in predicting user-item relationship,this study adopts the recommendation system model based on deep learning for drug repositioning research.We propose a neural collaborative filtering model named NCFBE which is based on deep learning technology combined with the latest data representation and integration methodsto detect potential drug-disease associations.This research is divided into the following two parts:(1)Firstly,the network representation method is appled to the biomedical knowledge base to learn the structure and semantic representation of biological entities automatically.Based on the learned vector representation of drug and diseases,the SVM is used to predict drug-disease associations.The experimental results show that the network representation of entities and the integration of multi-source data are beneficial to the predictive ability of the system.Further it shows a certain performance improvement compared with other systems in the same period.(2)Secondly,this study propose a novel framework named NCFBE,which Combines the neural recommendation system model with heterogeneous data representation and integration technology to predict drug-disease associations.This study uses the known drug-disease relationship to build a neural system filtering model.Furtherly drug and disease-related knowledge base information is applied to the model as auxiliary information which alleviates the cold start problem due to data sparseness.Experimental results from various datasets show the proposed model has better performance compared with otherstate-of-the-artsystems on multiple evaluation values such as AUC,AUPR,and F1.Moreover,we use NCFBE model to other biomedical associations prediction to verify its robustness.The experimental result in disease-mi RNA association detection confirms the superior performance and generalization ablility.In summary,the NCFBE proposed in this paper achieves excellent performance in biomedical entity relationship prediction,indicating that the deep neural network architecture has advantages in the feature representation learning and integration of biomedical entities.It provides a new view for computational drug repositioning.