Research On Chemical Drug Combination Recommendation

Combination therapy is an essential component of clinical intervention,which requires a comprehensive evaluation of the effectiveness,safety,and tolerability of drug combinations by considering individual patients’ conditions,administration methods,toxic side effects,drug interactions,and other factors.Irrational drug combinations will cause drug tolerance problems or even severe adverse reactions to patients.Traditional drug combination recommendation methods develop expert systems based on knowledge bases such as drug manuals and medical literature,which rely too much on artificially set rules and simple knowledge reasoning mechanisms,making it difficult to deal with complex and ambiguous real-world clinical information and unable to meet individualized drug recommendation for patients.This study leverages big medical data and various deep learning techniques to address this issue to achieve personalized drug combination recommendations by integrating heterogeneous pharmacological data.The research focuses on clinical medication recommendations and encompasses drug-drug interaction extraction,drug-target relationship prediction,and drug combination recommendation.A decision platform for recommending chemical drug combinations is designed and implemented.The main contributions of this thesis are as follows:1.Aiming at the problem that error propagation and practical application are too dependent on external features in the process of drug interaction extraction,a Multi-Task model with Multi-Granularity information(MTMG)is proposed.The objective is to leverage the correlation between drug-drug interactions and drug named entities by designing the drug-drug interaction extraction task as a sequence labeling task.Additionally,two sentence-level semantic granularity auxiliary tasks are designed to utilize the dataset’s inherent prior knowledge and provide training guidance for the main tasks.The experimental results show that MTMG achieves an F1 value of 0.925 for the drug entity recognition task,0.777 for the DDI extraction task,and 0.851 for the overall F1 value.2.Aiming at the problem that the existing drug-target relationship prediction methods do not fully use the structural characteristics of meta-graphs in heterogeneous graphs,a Meta Graph-based method for predicting Drug-Target Interaction(MGDTI)is proposed to enhance the representation of drug-target node features.MGDTI obtains the interrelated knowledge of drugs,proteins,and diseases from external multi-source medical big data,and constructs a heterogeneous knowledge information network diagram.Through the metagraph-assisted learning method,the inter-node relationship and semantic features contained in the heterogeneous information graph are obtained,and the prediction model of the relationship between drugs and targets is established.The experimental results show that MGDTI achieves an AUPR value of 0.9417,an AUROC value of 0.9338,and an F1 value of 0.820 for drug-target relationship prediction.Additionally,the precision for the top 128 link predictions reaches 0.9688,and for the top 256,it reaches 0.9414.3.Aiming at the problem that the existing medication combination recommendation methods based on electronic health records do not fully use the information of multiple drug interactions,a medication combination recommendation model using the Temporal Attention Mechanism and Multi-Data Fusion(TAMMDF)is proposed.The recurrent neural network based on the temporal attention mechanism is used to extract the individual characteristics of patients.Combining multiple pharmacological network features enriches external drug knowledge to reduce drug-drug interaction rates.The experimental results show that the Jaccard similarity of TAMMDF is 0.5267,the F1 value is 0.6813,and the PRAUC value is 0.7778,and the DDI rate is 0.0643.4.Based on Spring Boot,My Batis,Vue.js,and B/S architecture,the chemical drug combination recommendation platform is designed and implemented using Java,HTML,and Java Script language.The data persistence is carried out through the My SQL database to realize the process and information management of drug and target data.The platform has the functions of user management,system management,drug target data management,pharmacological property mining,and drug combination recommendation.