Research On Drug-Drug Adverse Reaction Representation Learning Method Based On Heterogeneous Polar Network

Combination medication is a routine and common treatment strategy in disease treatment,and usually achieves better therapeutic effects.Therefore,drug-drug interaction(DDI)caused by combination medication has always been a hot research spot.In fact,DDI is a double-edged sword,with both positive and negative DDIs.Compared with positive DDIs,negative DDIs(called Drug-Drug Adverse Reactions,DDADRs)are relatively difficult to be found.Although a large number of in vivo and in vitro experiments are performed in the drug development stage to find out drug combinations that may lead to DDADRs,this process consumes huge costs and requires a long experimental period.With the development of information technology,a large amount of medical data has been aggregated and mined.Researchers began to predict DDADRs with machine learning methods.Although many effective DDADR prediction methods have been proposed,there are still many challenges,mainly including:(1)Most of the existing DDADR prediction methods simply use drug-related information(such as chemical molecular structure,protein,etc.)to obtain the feature representation of drugs.Because the biomedical information is not fully utilized,the obtained drug features cannot represent the rich semantic relationship between drugs.Besides,most methods are based on a single view.How to use multi-view embedding methods integrating different biomedical views to obtain the drug feature representations with richer semantic information is a very challenging problem;(2)Most of the existing DDADR prediction methods rely too much on labeled data,while the artificial labeling of medical data is difficult.The cost is extremely high.How to build a DDADR prediction model based on contrastive learning to reduce the dependence on labeled data is also a problem worth studying;(3)Because the DDADR prediction model based on contrastive learning is limited by its own training mode,the obtained drug feature representation has a poor quality.How to improve the DDADR prediction model based on contrastive learning to obtain the drug feature representation with good quality is a question worth exploring.In response to the above problems,we study the DDADR prediction problem and key technologies,and further propose the DDADR representation learning methods based on heterogeneous polar network.The main innovations of this thesis are as follows:(1)Proposing a DDADR prediction,MS-ADR,based on the heterogeneous polar convolutional network to solve the problem that drug features contain insufficient drug semantic information and the single-view based prediction models are not effective.Firstly,the drug heterogeneous polar network is constructed with the multi-source biomedical information related to drugs and known DDADR information,and the structural equilibrium theory is applied in the drug heterogeneous polar network for sign propagation to further enrich the semantic information between drugs;secondly,graph convolutional coding is performed on the drug heterogeneous polar network to obtain drug feature representations containing high-order semantic information;finally,a novel deep learning-based feature fusion framework is constructed to embed different biomedical view information into encoding structure to obtain multi-source drug feature representations for DDADR prediction tasks.The MS-ADR method forms the basis for subsequent research works.(2)Proposing a DDADR prediction method,DMVDGI,based on the contrastive representation of heterogeneous polar network to solve the problem of the existing DDADR prediction method’s over-reliance on labeled data.Firstly,the graph convolutional encoder is used to obtain the drug positive local feature representation and the drug negative local feature representation,and the global encoder is used to obtain the drug global representation;secondly,the model is continuously optimized by maximizing the mutual information between the global representation and the positive local representation while minimizing the mutual information between the global representation and the negative local representation to obtain the optimal drug feature representation;finally,the drug feature representations in different biomedical views are fused through a semantic-level attention mechanism to obtain multi-source drug feature representations for the DDADR prediction task.Compared with the MS-ADR model(achievement 1),the DMVDGI model reduces the dependence on labeled data.The DMVDGI method forms the basis of subsequent optimization studies.(3)Proposing a DDADR prediction method,SCLADR,based on the supervised contrastive representation of heterogeneous polar network to address the low quality of drug feature representations obtained by DDADR prediction models based on contrastive learning.Firstly,a supervised contrastive learning method was used to fine-tune the DMVDGI model(achievement 2),using sparse labeled data,adding cross-entropy loss to train the model;secondly,the implicit neural network(Implicit Graph Neural Networks,IGNN)is used as the encoder to encode the multi-semantic feature representation of the drug,so that the obtained drug feature representation is global;finally,during the training process,negative samples are generated in an end-to-end trainable manner rather than in a fixed manner.These improvements significantly improve the quality of drug feature representation,thereby improving DDADR prediction accuracy.We conduct extensive experiments on the Decagon model dataset,Drug Bank,SIDER and OFFSIDES datasets.Besides,we evaluate models with multiple metrics.The experimental results show that our proposed DDADR prediction models work better than some of the state-of-the-art models,which effectively realizes the prediction of DDADR,thereby helping doctors to formulate treatment plans and improving the efficiency of new drug research.