Research On Drug Combination Prediction Based On Deep Learning

Drug combination plays an important role in the treatment of many complex diseases,and it can promote the healthy recovery of patients from multiple therapeutic directions.Compared with a single drug,the drug combination can improve the therapeutic effect,as well as reduce the side effect caused by the increase of a single drug dose.However,it will consume a lot of human and material resources to verify the effectiveness of thousands of drug combinations by the clinical laboratory.The development of deep learning technology provides a new method for the mining and application of clinical medical data.Deep learning technology can be used to analyze the clinical medical data and screen out potentially effective drug combinations,which will greatly reduce the work of clinical experts and shorten the time to find effective drug combinations.This paper focuses on how to use the deep learning technology to improve the predictive accuracy of drug combinations,which mainly contains the following two aspects.(1)Concerning the problem of the insufficient extraction of the temporal sequence feature of patients and the medication knowledge,a temporal attention mechanism is proposed to extract the important temporal sequence feature of patients and the graph neural network is used to calculate the drug embedding representation vector integrated with the drug knowledge,so as to predict the drug combination accurately.First,a temporal attention mechanism is designed based on the recurrent neural network,which is used to obtain the temporal sequence features at different time steps.Such an attention mechanism consists of a two-layer recurrent neural network.On each layer,the attention weights are calculated from different views of the medical sequence data to distinguish the importance of the patient’s features at different time steps.The temporal attention mechanism can fully extract the important features of patients hidden in the medical sequence data,and store them in the form of key-value pairs.Then,drug combinations are converted into drug networks,where each drug node is connected to its combined drug through an edge.The embedding vector of each drug node contains its own attributes and neighbor features.The graph neural network is used to extract the knowledge of the association among drugs and integrate such knowledge into the lowdimensional representation vector of the drug node so that it can fully obtain the information of the entire drug network.Finally,the historical features of patients are retrieved from the memory storage mechanism of key-value pairs to calculate the output vector of historical features,which is combined with the feature representation vector of patients to predict the drug combination.The experimental results show that the temporal sequence features of patients and drug knowledge can accurately predict drug combinations,as well as control the probability of adverse drug reactions to a certain extent.(2)As for the problem of the complex calculation of the prior medical knowledge and the global feature of patients,a prior medical knowledge extraction rule is designed to calculate heuristic drug features,and the multi-head attention mechanism is used to obtain the global features of patients from different feature subspaces,which improve the predictive accuracy of the algorithmic.First,a constrained heuristic drug computation rule is designed to convert the complex prior medical knowledge into numerical values and establish the mapping relationship between the diagnosis and the drug.The prior medical knowledge provides heuristic drug features for the drug combination prediction,and it is helpful for the fitting of deep learning training parameters.Then,the multi-head attention mechanism is used to obtain the interaction relationships within different views of the medical data to extract the global feature of patients from the feature subspaces of different views of the medical data.Based on the self-attention mechanism,a query vector,a key vector,and a value vector are defined on each self-attention layer.Therefore,each self-attention layer can learn the importance of different elements in the medical sequence data,and the features from different attention layers are connected to obtain the global information from different feature subspaces.Finally,the memory storage mechanism of key-value pairs is used to store and retrieve the historical feature of patients to predict drug combinations.The experimental results show that the prior medical knowledge and the global features of patients can further improve the prediction accuracy,and the case analysis shows that the algorithm has a pretty well performance when analyzing the treatment of uncommon diseases.