Prediction Methods Of Drug-target Interaction Based On Multi-level Feature Extraction

Drug-target interaction prediction can effectively predict targets of a drug,or predict effective drugs for potential targets,which is a very important step in drug design.According to the prediction results of drug-target interaction,drugs can be designed to regulate the physiological function of the target,so as to achieve the goal of disease treatment.Furthermore,it can more efficiently determine the on target and off target effects of drugs,detect the effects and side effects of drugs in advance,and avoid useless work in the drug design process.In recent years,researchers have proposed more and more methods for predicting drug-target interactions.However,the data of drugtarget interactions are extremely limited,especially most drug-target interactions are concentrated on a small number of drugs and targets,resulting in insufficient training data for most drugs and targets.As a result,the prediction results of drug-target interaction still need to be greatly improved.In this paper,according to the characteristics of drug-target pair data,some research works on the drug-target interaction prediction have been carried out by using feature selection,classfication and deep neural networks.The main works in this thesis are as follows:(1)Drug-target interaction prediction via double classification strategiesThe number of interactions is limit,especially most interactions focus on only a few targets or a few drugs.This makes some targets have larger numbers of interactions(TLNI)while others only have smaller number of interactions(TSNI).Only using a classification strategy may not be able to deal with the above two cases at the same time.To overcome the above problem,a drug-target interaction prediction method based on multiple classification strategies(MCSDTI)is proposed.In MCSDTI,different classification strategies are respectively designed for TLNI and TSNI to predict the interaction.For TLNI,the classifier-based method is used,so as to make use of the feature represented by the attribute of the sample,which can use more information.For TSNI,similarity method is used to predict the interactions.The similarity method only uses similarity features that are higher layer features between samples,so that the prediction model can be better learned in the case of limiting positive samples.As a result,the advantages of these classification strategies can be better utilized by using different levels of features for different data.(2)Drug-target interaction prediction via multiple output deep neural networkMCSDTI can give full play to multi-level feature,but this method uses traditional methods,which are simple models and have many limitations.For deep learning,complex classification models can be automatically learned from data just by defining learning objectives.However,due to the serious shortage of drug-target pair training data,it is difficult to fully train the parameters of deep neural network.As a result,the existing deep neural network structure has insufficient capability of extracting features and discriminating classifications in the prediction of drug-target interaction.Therefore,a multi-output deep neural network(MODNN)method to predict drug-target interaction is proposed.This method improves the prediction effectiveness of drug-target interaction by using a newly designed auxiliary classifier.Multiple auxiliary classifiers are distributed in the bottom layer,middle layer and high layer of the neural network,so that the multi-output deep neural network has the following features: Each auxiliary classifier contains a loss function,and multi-layer loss function can increase the return gradient signal;The multi-level feature can be used in the training model at the same time and can be utilized in the unified framework.Multiple different classification loss functions are used to train the classifier so that the MODNN has a certain ability of ensemble classifier.(3)Drug-target interaction prediction via multi-subspace deep neural networks and graph auto encoderMODNN can simultaneously utilize multi-level feature,but it has not considered the situation when the feature dimension is high,the number of samples is small,the sample distribution is very sparse,and then there will be multiple models with similar prediction ability.There are deep learning models that can only learn one of them.However,according to human cognitive thinking,when one model is not far superior to the others,the results of these models should be integrated.Therefore,a new deep neural network with multiple subspaces is designed.The network is designed with a subspace layer and an ensemble layer.The subspace layer learns multiple different strong feature subsets through multiple neural network substructures.The ensemble layer can make comprehensive use of these strong feature subsets in a unified optimization framework to further learn the feature representation of drug-target pairs in the ensemble layer,so as to better predict drug-target interactions.The method simulates the process of human beings dealing with similar problems by first collecting the knowledge of others and then refining and processing.On the whole,multiple subspaces and graph self-coding deep neural network methods transform the difficult problem of sparse feature extraction into multi-level feature extraction,which reduces the difficulty of feature extraction at each level and can extract features more fully.In summary,three methods for predicting drug-target interaction are proposed in this paper.Drug-target interaction prediction via double classification strategies using different levels of feature to take the advantages of different classification strategies on different data.Multi-output deep neural network simultaneously uses multi-level feature to train neural network so as to make fully use of the parameter training of neural network.Multiple subspace deep neural network designs a multi-level feature extraction strategy,which reduces the difficulty of feature extraction at each level and can extract more features.The experimental results show that the model created by the method in this paper has a high degree of certainty and is effective for the prediction of drug-target interaction.