The Research Of Drug Repositioning Based On Deep Learning Methods

Drug design using machine learning methods has been one of the important methods in drug research.In recent years,with the successful application of deep learning in various fields,it is a subject of great scientific and practical significance that how to apply the framework of deep learning to drug repositioning.In the thesis,we use the methods combining all sorts of deep structures with drug’s attribute information and relationship to mine the patterns about drug repositioning,then we will expect to find novel indications for some drug canidates,thus our methods could help drug researchers to accelerate the research speed of drug repositioning.The research work of this thesis includes the following four aspects:（1）The recommendation algorithm for active small molecule is studied and the recommendation algorithm is based on the activity information of drugs and targets,the feature information of targets and the structures information of drugs.In this part,the UserbasedCF is used for the recommendation task,and then AEuserbasedCF is proposed to improve the results.In the AEuserbasedCF model,the AE model is used to reduce the dimension of the feature information of targets,after that the dimension reduction data is imported the UserbasedCF model to compute the recommendation results.From the final results,it can be proved that the AEuserbasedCF model can improve the accuracy of the recommendation.Then,the HybridSimCF model is proposed to solve the cold start problem in recommendation system.According to the commonly used ligand-based method in this field,the 2D structure information of some drug molecules are extracted and are preprocessed to obtain a similarity matrix of small drug molecules.By comparing with other algorithms,the HybridSimCF model can effectively solve the cold start problem and improve the accuracy of the recommendation task.（2）The research of deep learning algorithm is based on the relationship between drugs and diseases.In this part the drugs and the diseases are mapped to two different sets in a bipartitet network,through this transformation,the drug repositioning problem is converted into a link prediction problem of complex networks,and then the Restricted Boltzmann Machine（RBM）model is applied on this dataset.From the AUC values,we find that the RBM model has an advantage in mining potential patterns of the relationship between drugs and diseases,the prediction accuracy is better than other algorithms,and some candidate drugs in the prediction results are also proved in the relevant database and literature.Then we use the idea of the link prediction in complex networks which is the more similar the nodes are,the more likely they are to generate links,the features which can describe the similarity of nodes are extracted,all of the features are composed the tag data set,and the supervised learning of NN_RBM model is carried out on the data set in order to discover the possible links,and the new method is applied to speed up the training precess.（3）The recommendation algorithm is studied by adding the related properties of drugs,proteins and side effects to the relationship between drugs and diseases.Before using the UserbasedCF recommendation algorithm to recommend candidate drugs,the deep dimensionality reduction method is adopted for the added high-dimensional attributes,and the two different dimensionality reduction frameworks are used to reduce the attribute data dimensions,the two different structures are named by Deep_Framework₁ and Deep_Framework₂,and then the tranditional PCA is also used to the same data.The three reduction data are used for computing the similarity matrix and completing the recommendation task.By comparison,Deep_Framework₁ and Deep_Framework₂ can achieve good results,and Deep_Framework₂ has the best performance in the three methods.After that we have built a small query system that displays the drugs that can be recommended for each disease by query.4)The multitask DBN model is applied to the prediction of drug repositioning.In drug repositioning,one guiding idea is that when drugs are found to recommend for other disease,the fewer side effects are expected to be better.Since most drugs can cause side effects when they are applied to the human body,it is necessary to compare the side effects of the drug after finding new indications.This is obviously a multitask problem,and this part will try to solve this problem by using the multi-task deep learning model.All diseases are classified 20 different classes,and the drug’s side effect will be classified by the number.After that we will build a dataset that fits the DBN and then differentiate the importance of the task,adjust the dataset through the sample balance strategy.By comparing with MNN model and MDBN model,it is found that MDBN model is suitable for drug repositioning,which has a better and more stable effect in predicting drug candidates.Through the above four aspects,it is proved that the deep learning methods can be regarded as a good research tool used in drug repositioning.The deep learning methods can help the researchers speed up the discovery of drug condidates,and reduce the cost of development in drug repositioning.