Research On Feature Selection Method For Multi-Cell Line Drug-Target Interaction Prediction

Drug-target interaction(DTI)prediction plays an important role in drug development and application.By exploring new indications for existing drugs,we can speed up the progress,reduce the cost and risk of drug research and development.Traditional drug-target pair testing is expensive and tedious.And the development of sequencing technology in recent years has led to a dramatic increase in the number of potential features in drug and target research.All these affect the efficiency of drug research and development.In this dissertation,the prediction of DTI is defined as a binary classification problem,and the feature selection in the prediction process is studied by using machine learning method.It mainly includes three aspects:the representative of feature subset,the efficiency of feature selection process and the stability of feature selection method.Specific research contents are as follows:1.A feature selection algorithm based on Randomized Logistic Regression(RLR)and greedy strategy is proposed.This algorithm can minimize the dimensionality of data without affecting the accuracy of prediction,and select a set of effective feature subsets suitable for multiple classifiers to realize the prediction of DTI in different cell lines.2.The Siamese network framework is used to select and learn the features of the same gene locus of drug and target by the deep neural network(DNBN)with shared parameters.Then,the feature level fusion is used.Thus,better prediction results can be achieved in different cell lines with fewer parameters.3.Combining the ensemble learning idea with feature selection method,an ensemble feature selection method can be used to select effective feature subsets steadily from the sample sets of different cell lines,which is suitable for different classification algorithms to predict DTI.In this dissertation,the experiment bases on drug and target expression profile data from multi-cell line.The experimental results of three feature selection methods verify that:for different data sets and classifiers,our methods can select representative features with high efficiency and stability.At the same time,these studies play an important role in predicting DTI.