Prediction Of Drug Target Interactions Based On Multilayer Network Representation Learning

With the development of biotechnology,more and more human proteins are discovered by researchers.However,among the known proteins,the number of proteins that can be used as targets is scarce,accounting for only a small part of the total number of known proteins,resulting in the number of target proteins of most drugs being only two or three.If more potential drug target interactions can be discovered,more complex diseases can be treated,thereby reducing the time,cost and risk of new drug development and providing a safe guarantee for human development.When exploring an unknown drug target interaction,traditional thinking chooses to use experimental methods.Although the results of the experiment are reliable,but the time period is long,the cost is high,it is impossible to detect all possible drug target interactions.Therefore,it is necessary to use a computational method to narrow the possible detection range for experimental verification,reduce the experimental cost,and reduce the experimental time.At present,computational prediction drug target interaction algorithms are mainly divided into two categories based on data types: first,based on single-type data prediction;second,based on multi-type data integration prediction.However,none of these methods can deal well with the effects of noise in multi-dimensional data,and when extracting the characteristics of drugs and targets,the deep nonlinear topological structure relationship in the network is not captured.These factors will affect the final prediction accuracy.This paper proposes a multi-layer network representation learning for drug-target interactions prediction(MEDTI)based on multilayer network representation learning.Represent a learning algorithm through a multi-layer network,integrate multi-layer network information,learn the compact features of drugs and targets,and then use the geometric proximity of related vectors in the same space to use known drug target interactions as supervisory information to obtain drug-targets interactions prediction score matrix.After obtaining the drug target interaction score matrix,the algorithm performance analysis and result verification were performed on the prediction results.The MEDTI framework proposed in this paper integrates multiple types of data well,effectively reduces the impact of multi-dimensional noise in multi-layer networks and improves the accuracy of prediction by improving the learning method of multi-layer network representations for predicting drug target interactions.Provides a new perspective for drug-target interactions prediction.In the experiment,this paper compares the MEDEI algorithm with the DDR,LRSSL,MDA,and MNE methods in a unified data set.The result is that the MEDTI framework has the highest prediction accuracy.Experiments also prove that integrating multiple types of data helps to mine more hidden information in the network,and can improve the accuracy of predicting drug target interactions.In other experiments,the influence of the parameters of the algorithm on the stability of the algorithm is also analyzed.In the result verification,different verification methods were used to verify the prediction results,which improved the reliability of predicting the drug target interaction results and proved that the framework MEDTI can be used to predict drug target interactions.Exploring new drug target interactions for experiments narrows the scope of the experiment,predicts potential drug target interactions,reduces experiment costs,and shortens experiment cycles.