Research On Prediction Of Drug-Target Binding Affinity Based On Graph Neural Network

In recent years,people’s concern for health has continued to increase,and higher demands have been placed on pharmaceutical-related products.The binding affinity between drug and target,i.e.the strength of their interaction,is one of the most important research objectives in drug development.In related research,traditional experimental assays can show more accurate 3D processes but cannot meet the demand of R&D efficiency,and machine learning methods achieve a certain degree of accuracy and fast prediction at the expense of a large amount of implicit data information.As a result,scientists are eager to seek an accurate and efficient prediction method for technological innovation.Graph neural network is a method for modeling and analyzing graph structure data,which can show strong characterization and generalization ability for drug molecules with complex connectivity and large scale of point-edge data.This paper proposes a graph neural network model called DC-DGNN to convert drug molecule graph structures into low-dimensional vectors from molecular graph representations to address the problems of limited depth of graph convolution due to oversmoothing and slow computation of large-scale graph data.The method uses a graph convolutional structure which is densely connected to deepen the network,and proposes a new hierarchical attention mechanism pooling GApooling with the traditional global pooling to form a dual-channel pooling.GApooling can be embedded in each layer of the network to adaptively measure the importance of each substructure and eliminate unimportant information,thus reducing the graph size and computational effort.Combining the work of multi-channel convolutional neural network for protein sequence embedding representation of targets and coding processing of related data,we train the DC-DGNN graph model for prediction on two commonly used datasets,Davis and KIBA.The experiments show that the DC-DGNN model has improved in all metrics compared with recent machine learning,deep learning,and graph neural network baseline models,which corroborates the wide application prospects of graph neural network in the field of drug-target binding affinity studies.