Drug Discovery Prediction Based On Network Representation Learning

In recent years,the field of medicine has developed rapidly.As a challenging task in the field of medicine,drug discovery has always attracted the attention of researchers.Especially in the process of drug discovery the prediction of drug interactions and the prediction of drug-target protein interactions are of great significance to drug development.Generally speaking,The prediction of drug-target protein interactions in the process of drug discovery is the basis for determining a new drug molecule,as well as the basis for predicting drug interactions.The prediction of drug-drug interaction is the key to detecting whether a new drug has an interaction with other drugs,and it is also an important process to ensure the safety of new drug discovery.In the process of drug discovery,traditional laboratory methods are time-consuming,labor-intensive,and blind.With the development of computer technology and the continuous accumulation of medical data.Researchers continue to use algorithms in the computer field for drug discovery,which solves the shortcomings of traditional laboratory methods.For drug-target interaction prediction,methods based on machine learning and deep learning usually use the chemical structure information of drug molecules and target protein molecules and their rich external relationships to train prediction models,but in these methods,there is almost no attention to the different characteristics of the relationship between different topology and rarely consider the heterogeneous network topology information.Therefore,a heterogeneous network based on relational topology is proposed to predict the relationship between drug and target.First,the heterogeneous network is embedded in a low-dimensional space to learn the vector representation of the node to make full use of the network topology information.Then,according to the topological structure of the different relationships between nodes in the heterogeneous network,the relationship is divided into two types of affiliation relationship and peer relationship and modeled separately,so as to better capture the semantic information between different relationship nodes.In addition,the model proposed in this dissertation can also obtain good prediction results without adding drug-target pairs with known interactions into the training set,which solves the problem of low prediction accuracy caused by insufficient drug labeling data to a certain extent.In this dissertation,experiments were carried out on open drug data sets,and the predicted AUC value reached 96.93% compared with the baseline method in the case of high prediction efficiency.For the prediction of drug interactions,the methods based on computer have achieved some achievements,but there are some limitations in some aspects.For example,most of the methods do not consider the characteristics of drug network topology.In addition,most existing methods are based on information about drug properties,which is difficult to obtain before a new drug is released to the public,so it cannot be used to predict drug interactions in the drug discovery stage.Therefore,in response to the above problems,this dissertation proposes a hybrid text-aware network embedding method,MTNE,to predict whether there is an interaction between drugs.In this method,the easy-to-obtain drug description information and pharmacodynamic information are mainly used.In addition,the mutual attention mechanism is also used to learn the dynamic text representation of drugs,and the text-based representation and the network topology-based representation are effectively combined to enhance the performance of prediction.Experiments on the Drug Bank dataset demonstrate the effectiveness of the proposed method.In addition,compared with the existing algorithms,the proposed model improves the prediction performance by about 6% on a small-scale dataset.Finally,the drug discovery prediction method based on network representation learning proposed in this dissertation not only makes full use of the different relationship structure characteristics of drug-target network,but also makes use of the rich text information of the drug,which speeds up the process of drug discovery to a certain extent.