| Herbs and prescriptions are unique clinical intervention methods of traditional Chinese medicine,and they are also important treatment methods for individualized diagnosis and treatment of chronic and complex diseases in today’s society.The modernization of herb,with the core task of revealing the micro-pharmacological mechanism of prescriptions,has always been an important topic in traditional Chinese medicine research,and it plays an important role in promoting the development of the academic and clinical diagnosis and treatment of traditional Chinese medicine.However,due to the diversity and complexity of herb components(the components of a single herb),the pharmacological effects of herb often exert their effects through multi-target pathways.Therefore,using traditional pharmacological experimental methods to confirm the pharmacological mechanism of herb,especially the identification of herb targets,there are many challenges,which is an important bottleneck in the modernization of herb research.In addition,herbs,as the basis of clinical compound medicines in traditional Chinese medicine,often combines multi-flavored herbs systematically to form clinical prescriptions with overall disease control effects.The characteristics of multi-component,multi-target and multi-step integrated synergistic effects of prescriptions have brought difficulties and challenges to the study of the mechanism of action of herb compounds.Therefore,there is an urgent need for new technologies and new methods suitable for the study of the mechanism of action of herbs,design effective herb target prediction models,explore the molecular mechanism of prescriptions,and reduce the time and capital cost of herb research and development.On the other hand,the successful application of network medicine and network pharmacology shows that studying the mechanism of action of herbs from the perspective of the network is an effective way.Therefore,this dissertation proposes four methods for predicting herb targets based on herb clinical theories,and further extends the evaluation of prescriptions based on the relationship of herb-target.The specific research of this article is mainly divided into the following four parts:(1)Aiming at the problems encountered in the construction of networks and feature learning in traditional herb pharmacology research.This chapter makes full use of the knowledge of the medicinal properties and clinical effects of herbs,and combines the existing genome and disease mechanism data.Explored the feasibility of integrating heterogeneous networks to predict the relationship between herbs targets from a clinical perspective.A prediction framework HTINet based on the knowledge of the medicinal properties and clinical effects of herb is proposed.With the help of the network representation learning method,there is no need to consider the chemical composition of herbs and the molecular information of the targets,combined with the topological characteristics of the nodes in the multi-layer heterogeneous network,learn the lowdimensional dense feature representation of the herbs and the targets,and build a classification model prediction.The experimental results show that HTINet has better performance than other baseline methods,and also shows the feasibility of predicting the interaction of herb targets from a clinical perspective.(2)Aiming at the traditional network-based prediction method of herb targets,there is no problem of co-optimization combined with subsequent target prediction tasks during node feature learning.This chapter builds an end-to-end prediction model,which combines the learning of the characteristic representation of herbs and the targets on the network with the subsequent target prediction task,and the latter is used to further optimize the characteristic representation.And based on the graph neural network,through the application of multi-layer aggregation to the local representation to capture higher-order information,finally two herb target prediction methods based on the graph neural network are proposed,HTIGCN and HTIGAE.HTIGCN mainly obtains the lowdimensional eigenvectors of herbs and targets through graph convolutional networks,combines the topological properties and neighborhood information of each herb(or target)node in the network,captures nonlinear interactions,and uses matrix compensation train and optimize the feature learning of nodes.HTIGAE learns the low-dimensional feature representations of herbs and targets through graph convolutional networks.On the one hand,the feature representations of the two are mapped to the same space for supervised learning.On the other hand,the original input network is reconstructed based on the graph autoencoder,and the feature representation of the node is optimized by calculating the global network structure error.Experimental results show that the two models outperform other baseline methods on most metrics.Finally,we manually verified some of the predicted target interactions of herbs based on biomedical literature.(3)Aiming at the problem of the lack of consideration of the synergy between multiple tasks in the existing prediction methods of herb targets.We designed the prediction framework HTIMTL based on multi-task learning.By introducing high-quality data of traditional herb pharmacology as supervised information,we implemented auxiliary information sharing and embedding,and learned the embedding of joint multiple network tasks for each node,through an end-to-end approach more effective training for the underlying herb target prediction task.Experimental results show that the predictive model is superior to other baseline methods in most metrics.(4)In view of the current lack of appropriate methods to evaluate the internal mechanism of multi-flavored herb combinations(herb compatibility)in prescriptions.Based on the target relationship of herbs,combined with genomic data,we use complex networks to study the evaluation methods of prescriptions.Starting from a large-scale prescription dataset,we constructed a weighted herb combination network.Through network characteristic evaluation,it is found that there are hierarchical characteristics in herb combination network.Then,based on the target relationship of herbs,a method for evaluating the molecular network association of herb compatibility was designed to quantitatively evaluate the prescription.Finally,the correlation between the quantitative scoring of prescriptions and the stratification characteristics is discussed.The results show that the higher the degree of stratification of the prescription,the better the effect of quantitative scoring.This result also reflects the generally recognized principle of "junior,minister and assistant" in TCM theory,and provides reference for the clinical application of combined treatment of chronic diseases and the further development of multi-drug combination. |
