Research On Drug Side Effect Association Prediction Method Based On Deep Learnin

Studying drug and side effect association prediction can help scientists better understand the mechanism of action and possible side effects of drugs in the human body,thus improving the success and safety of new drug development.Side effects are one of the main reasons for many drug development failures,and if computer models can be used to quickly and accurately screen out potential side effects related to drugs,a lot of laboratory testing costs can be saved,thus reducing the time and cost of new drug development.In order to better understand and utilize the effects of drug treatment,scientists have been trying to predict the relationship between drugs and side effects through technological means such as big data and artificial intelligence.However,many methods do not sufficiently learn and incorporate information from different perspectives that might be useful for forecasting.In this paper,three computational models for predicting the association between drugs and side effects are proposed based on heterogeneous infographics composed of multi-source data related to drugs and side effects.CCRS is a predictive model based on attribution-level attention convolutional neural networks to learn and deeply integrate attribute information about pairs of drugs and side effects in multiple heterogeneous graphs.Multiple drug-side-effect heterogeneous graphs are established to learn the paired attribute information of drug-side-effect nodes,which can reflect the similarity of drugs from different perspectives.And multiple attribute embeddings of drug-side pairs are extracted from multiple heterogeneous graphs.Multiple enhanced attribute embeddedness is obtained based on the attribute level attention mechanism,which captures the different importance of each attribute within each attribute embeddedness to association prediction.Multiple paired attribute representations are obtained based on multi-layer convolutional neural networks to deeply integrate a pair of drug and side effect node related attributes.Finally,multiple paired attributes were represented by fusion and the predrug-side effect association score was performed.The prediction ability of CCRS is demonstrated by several experimental evaluations and case studies.The topology of heterogeneous graph contains the relationship and connection mode between nodes,reflecting the structural characteristics between drugs and side effects,which is important for the prediction task.Therefore,a drug-side effect association prediction model based on graph convolutional autoencoder,GCRS,is proposed.Firstly,the same strategy as CCRS was adopted to construct multiple double-layer heterogeneous graphs based on the similarity of multiple drugs,which were used to embed the network topology information and node attribute information of drug and side effect nodes.Since each heterogeneous graph has its specific topology,a unique module based on the graph convolutional autoencoder(GCA)is built to learn the specific topological representation of each drug node and each side effect node separately.Because multiple graphs reflect the relationship between drug and side effect nodes,they contain a common topology.Therefore,a GCA module based on shared parameters is also constructed to learn the common topological representation of each node.The attention mechanism at the topological representation level is designed to obtain a more informative topological representation.Finally,the network topology representation and paired attribute representation of nodes are fused to obtain a pair of drug and side effect association scores.The experimental results show that GCRS has superior performance.The semantic relation between nodes reflects the semantic relation and interaction between nodes,so that more knowledge and information can be inferred and predicted more accurately.In order to fully understand the semantic relationship between nodes,a drug-side effect correlation prediction model based on collaborative knowledge distillation,KGCRS,was proposed.Metapath is a graph-based method used to extract semantic information between nodes.It defines a path pattern based on node type and edge type to describe the relationship between nodes.Therefore,this method adds a method based on collaborative knowledge distillation to learn the semantic information of nodes from the metapath.Due to the different information contained in different heterogeneous graphs,we also designed the attention mechanism of semantic representation level and attribute representation level to obtain more informative semantic representation and attribute representation respectively.Then,fully connected neural networks were used to integrate paired attribute information,unique and shared topological information and semantic information to obtain drug-side effect association scores.The experimental results show that KGCRS has better performance and can predict more accurate correlation.