Research And Implementation Of Computational Drug Repositioning Based On Deep Learning

Computational drug repositioning,an important and efficient method of drug discovery,aims to identify new uses for known drugs.Drugs that have been approved by drug administration have good performance on bioavailability,safety and pharmacological characteristics,thus the technology of computational drug repositioning can significantly accelerate the drug discovery process,reduce investment costs and strengthen drug controllability,which has more advantages compared with the traditional process of new drug discovery.At present,an increasing number of researchers use the idea of collaborative filtering model to solve the problem of computational drug repositioning.However,there exist several problems in the previous research: Firstly,the sparseness of the computational drug repositioning data makes the related models vulnerable to cold-start problems,thus the inference ability of these models would be limited;Secondly,the matrix factorization model in previous studies has insufficient learning ability,thus it is unable to learn higher-level drug-disease associations;Furthermore,the previous models do not take the effect of negative samples into consideration,which could lead to over-fitting phenomenon and thus limits the generalization ability of the models.Finally,although the computational drug repositioning based on latent factor model can effectively capture the overall information shared by a majority of drug-disease associations,it is difficult to mine the neighborhood information contained in a small number of strong drugdisease associations.Hence,to overcome the above problems,this thesis uses matrix factorization,deep learning and other related technologies to propose the computational drug repositioning model(Additional Neural Matrix Factorization,ANMF)based on matrix factorization and neural network,and the computational drug repositioning model(Hybrid Attentional Memory Network,HAMN)based on memory network and attention mechanism.The main research work of this thesis is as follows:(1)In terms of the sparseness of the data,insufficient learning ability and lack of consideration on the role of negative samples of the previous models,this thesis proposes a computational drug repositioning model Additional Neural Matrix Factorization(ANMF).Firstly,an improved autoencoder is used by this model to extract the hidden features of drugs and diseases.At the same time,the drugs-drugs similarity and diseases-diseases similarity are included in the process to enrich the characterization information of drugs and diseases to overcome the problem of data sparseness.Then the ANMF model uses neural networks to replace the inner product operation to generate predicted values.The introduction of neuron nodes enables the model to consider the weight relationship between hidden features.In the meantime,the combination of non-linear activation functions can help the model learn more complex associations between drugs and diseases to improve the learning ability of the model;Finally,the AMNF model uses negative sampling technology,allowing the model to take into account the role of negative samples to reduce the risk of over-fitting.The effectiveness and practicability of the ANMF model are demonstrated by the experimental results on two real computational drug repositioning datasets.(2)The ANMF model,in essence,is a computational drug repositioning model based on latent factor,which could effectively capture the overall information shared by a majority of drugdisease associations.In order to take into account the neighborhood information contained in a small number of strong drug-disease associations,this thesis introduces a computational drug repositioning model Hybrid Attentional Memory Network(HAMN).The main idea of this model is to use the attention mechanism and memory network to generate the neighborhood model,and then incorporate the neighborhood information contained in some strong associations learned by the neighborhood model into the ANMF model.Specific steps are as follows.Firstly,an improved autoencoder is used by the HAMN model to capture the overall information shared by a majority of drug-disease associations.Then,the HAMN model combines the attention mechanism with an external memory unit to generate the neighborhood contribution representation,which is used as the neighborhood model to capture the neighborhood information contained in a small number of strong drug-disease associations;Finally,the HAMN model uses a non-linear function to integrate the overall information shared by a majority of drug-disease relationships with the neighborhood information contained in a small number of strong associations to obtain a predicted value.The related experimental results on two real computational drug repositioning datasets show that including neighborhood information into the latent factor model can effectively improve the model’s performance and generalization ability.