Network Pharmacology Analysis And Application Based On Deep Neural Network And Binding Modes

With the development of modern rational drug design techniques and the accumulation of big data,the traditional 'one drug,one target,one disease' hypothesis and the research regime based on this hypothesis is showing their weakness in efficient drug discovery.Since the idea of the correlation between biological networks and Traditional Chinese Medicine(TCM)was firstly proposed,and the concept of 'drug targeting a network target instead of a single target' has been arousing interests and been widely applied in the process of the modernization of TCM.By constructing network motifs related to a particular phenotype or disease(also known as the 'network target'),mapping the drug targets on the network,and the subsequent enrichment analysis and topological or dynamical analysis of the network,researchers could reveal the biological basis of complex diseases on the level of networks,identify the key components in the network,as well as discover the active candidates among a group of compounds,especially components originated from the herbal formulae.While the concept of the 'network target' and the subsequent computational utilities has been developed and successfully applied in practice,there are difficulties and challenges in this area.Firstly,the boundary of the actual network motif consisted by biological entities are rather vague,lacking solid validation mechanisms.Secondly,current procedures of network pharmacology analysis don't involve further lead optimization,thus fail to facilitate the next-step development of drug candidates.Furthermore,current network pharmacology analysis usually stops at the inexplicit mapping of predicted drug targets on the network,which is weak in forming pharmacological hypothesis.In this study,a novel research regime based on the network of 'phenotype-protein targets-the binding mode between protein targets and ligands' is proposed,featuring the construction and training of the multitask neural network QSAR model,and the visualized structure-activity relationship(VISAR).The workflow of VISAR starts from the target sets related to a specific phenotype,and utilizes the rich compound-protein interaction datasets to build and train predictive models(especially multitask neural network QSAR models).Next,by dissecting the chemical features learned by QSAR models,the transformed chemical features of the neural network hidden layers and the derivative of features are used for activity landscape and atom contribution analysis respectively.The landscape and atom contributions are explainable and quantitative interpretation of the binding modes,and critical in understanding why the models give certain prediction from global and local perspective.Finally,the correlation between targets and binding modes and the correlation between phenotype and binding modes could be inferred and further explored based on the landscape and atom contributions,thus generating hypothesis on the biological basis of network targets and novel strategies for the design of drug combinations and lead optimization.In summary,VISAR promoted a novel way in understanding the network targets,dissected the underlying mechanism of network targets from a different perspective,and proposed a new research regime for network pharmacology.The major discoveries and innovations in this study are:(1)the first development of an interactive tool for dissecting chemical features learned by neural network QSAR models,which becomes the basis for the subsequent analysis of the binding modes;(2)the preliminary validation of the VISAR workflow by the case studies of GPCR and kinase family inhibitors;(3)the first application of VISAR on the analysis of immunogenic cell death(ICD)inducers,which identified a novel binding mode that correlated both with ICD phenotypes and cold and hot nature of TCM,and proposed a series of candidate chemicals originated from TCM herbs with hot nature,which could be further validated by experiments.