The Research On Drug Target Protein Prediction Algorithm Based On Semi-supervised Learning

Finding the interaction between drugs and targets is a key point in drug research.Altho ugh those in vitro methods for drug-target interaction predicting are all based on the powerful phar macological theory,those methods are all time co nsuming and costly which becomes the bottleneck of drug research and restricts their application.So a set of co mputational methods are needed to give more useful information to in vitro methods.The methods based on machine learning theory are diverse and also the focus of the drug-target prediction research.Among those methods semi-supervised learning prediction methods have attracted more attention for their advantages of adopting a large mo unt of unknown drug-target interactions.Finding an efficient se mi-supervised learning prediction method is a meaningful and challenging task.The main conte nt of this paper is described as follow:The drug-target interaction matrix adopted by original se mi-supervised learning prediction model is too sparse and imbalanced also with a lot of false negative data.To overcome the problems talked above,a new label extended method is propsed.This method first extends the original label matrix and then introduces a weight matrix into the lost function of the prediction model.To reflect the real similarity between each drug we then use the ATC code of drugs constructing a new drug similarity matrix and this matix has been used by the label exte nded method.The two kinds of cross validation e xperime ntal results show that the new label extended method achieve higher ROC-AUC and AU-PR with more efficient prediction perfor mance co mpared with ma ny prediction methods and at the sa me ti me ma ny new drug-target interaction pairs have been predicted on the four data sets by this new prediction method.The way of measuring the similarity is the key to all the si milarity based se mi-supervised learning prediction methods,however,the tradictional way of measuring the similarity has a limitation for it can no t capture the different impota nce of each small structure of drugs or tagets in drug-target interaction prediction.Faced with this problem a new feature vector based muti-label learning method is proposed.This method first adopts the semi-supervised feature selection algorithm to extract the main features of the original structure infor mation and the n use the muti-label learning model to learn the weight of each s mall structure giving the linear prediction function a nd finally for m the prediction result.The result of experi ment reveals that the prediction method of treating each small structure differently can really achieve better prediction perfor mance compared with those prediction me thods a dopting tradicional similarity measurement.