Research On Adverse Drug Reaction Based On Machine Learning

With the widespread uses of drugs,the issue of drug safety has attracted considerable attention in recent years.As one of the main concerns in the field of drug safety,adverse drug reaction?ADR?has become the fourth leading cause of death following cardiovascular disease,cancer,and infectious diseases.Specifically,ADR is estimated to result in as many as 770,000 injuries and death annually around the world,which not only leads to a tremendous economic pressure in patents,but also wipes about a significant budget in healthcare systems.Incompatibility of Traditional Chinese medicine?TCM?and adverse drug-drug interaction?ADDI?are primary issues in the research of ADR.By developing machine learning models,this thesis focuses on formulating the incompatible relationship among herbs to explore the potential incompatible herb combinations and modeling the adverse interaction among drugs to predict potential ADDIs.The main research work of this thesis is presented as follows:?1?Anti-community detection in complex networks is capable of exploring the negative relations among objects.Meanwhile,incompatible herb combinations can be seen as negative relations among herbs,which are always with low co-occurrence frequency in the same prescriptions.By first constructing herb network with high co-occurrence frequency of herbs in the same prescriptions and then detecting anti-community structure in herb network,one can explore incompatible herb combinations inside anti-communities.To formulate anti-community structure in complex networks,this thesis analyzes the effects of degree and self-edge of nodes on stochastic block model,and proposes a Degree stochastic Block Model?DBM?and a No s Elf-edge Stochastic bl Ock Model?NESOM?for detecting anti-community structure.Based on the formulations of the proposed two stochastic block models,two objective functions L?C?and Q?C?for the evaluation of anti-community structure,two synthetic benchmarks DBM-Net?DBM-Network?and NESOM-Net?NESOM-Network?for the validation of the performance of anti-community detection algorithms,as well as two algorithms LEOA?Local Expansion Optimization anti-community Algorithm?and NESOM-AC?No s Elf-edge Stochastic bl Ock Model Anti-Community algorithm?for anti-community detection are properly introduced,respectively.In particular,by developing a balance factor and designing a balance strategy,NESOM-AC can detect more balance anti-community structure and explore more negative relations among objects than the compared methods.Extensive experiments on synthetic benchmarks and real-world networks demonstrate the effectiveness of the proposed algorithms in anti-community detection.?2?For exploring incompatible herb combinations in TCM,this thesis proposes an anti-community detection algorithm based on Random non-n Eighboring n Ode expansio N?REON?and an ANti-community detection algorithm based on the DEgree and the RATio between the Inner degree and the Outer degree of a Node?ANDERATION?.By calculating the shortest distances among nodes,REON detects the node set by non-neighboring node expansion and regards the node set with the highest anti-modularity as an anti-community.By analyzing the effects of node DEGree?DEG?and Ratio between the Inner degree and the Outer degree of a node?RIO?on anti-community detection and finding that the nodes with high DEG and low RIO locate more fixedly in the anti-communities than the ones with low DEG and high RIO,ANDERATION preferentially considers the node with high DEG and low RIO for heuristic anti-community detection.Experimental results demonstrate that the proposed algorithms are effective in detecting anti-community structure in the herb network.Through the selection and verification by TCM experts,it can be found that the potential incompatible herb combinations can be explored inside the detected anti-communities.?3?To better predict incompatible herb combinations in TCM,this thesis proposes a non-negative matrix tri-factorization based learning model IHPreten?Incompatible Herb combinations Prediction with herb attributes and their correlation?by employing two herb attributes?efficacy and flavor?and their correlation to formulate the incompatible relationship among herbs.In IHPreten,a hypothetical test is conducted to evaluate the statistical significance of the dissimilar characteristics of efficacy and flavor in incompatible herb combinations and the attribute information from TCM literature is introduced for attribute supervision.These two constraints are jointly incorporated as regularizations into IHPreten for modeling the incompatible relationship among herbs.Experimental results on the real-world cold,cool,mild,warm,and hot incompatible herb combination datasets demonstrate the encouraging performance of IHPreten in predicting incompatible herb combinations in TCM.?4?In the research of ADDI prediction,this thesis proposes a Multi-Task Multi-Attribute?MTMA?learning model and a Dependence Guided Discriminative Feature Selection?DGDFS?model,which incorporate two drug attributes,molecular structure and side effect to model the adverse interactions among drugs and introduce self-representations of drug attributes for the linear reconstruction of molecular structure and side effect matrices of drugs.Meanwhile,to explore the leading factors in ADDIs,MTMA model imposes l_2,1-norm on the predicted molecular structure and side effect matrices of drugs,while DGDFS model employs discriminative feature selection and introduces l_2,0-norm equality constraints to molecular structure and side effect matrices of drugs for selecting discriminative molecular substructures and side effects accounting in ADDIs.Experimental results on the real-world dataset demonstrate the considerable results of the proposed models in predicting potential ADDIs and exploring the leading molecular substructures and side effects in ADDIs.