Drug-target Affinity Prediction Based On Multi-head Self-attention And Multi-channel Graph Convolution

Computer-aided drug design with high performance is a promising field,and the prediction of drug-target affinity is an important part of computer-aided drug design.It can give information on the interaction strength between the drug molecule and the target protein,to judge the degree of binding between small molecules and proteins.The computational method can quickly and accurately predict the affinity value between the target and the drug,which not only shortens the drug development cycle,but also saves the drug development cost.In recent years,various computational methods for predicting drug-target affinity values have emerged,such as molecular docking,ligand-based methods,machine learning and deep learning.In the above methods,deep learning,as a special type of machine learning,not only simplifies the steps of data feature extraction,but also increases the prediction accuracy of the model algorithm.Graph neural network series algorithm has been gradually applied in drug and protein research due to its excellent performance in structural feature learning.In the new field of drug-target affinity prediction,graph neural network also has great potential.Attention mechanism can obtain the relationship between global and local features in one step.At the same time,it is designed into parallel mode to reduce training time.It is widely used in deep learning research.In the new field of drug-target affinity prediction,graph neural network and attention mechanism have enormous potential.Firstly,a method of drug-target affinity prediction based on multi-channel graph convolution network is proposed in this paper.This method extracts the physical and chemical characteristics of drug molecules and protein sequences and the correlation map data as the model input,in which the map data corresponds to the node adjacency matrix of drug molecule map and protein target map.Then,a multi-channel graph convolution network is constructed,which aggregates the information of different distance nodes.Through the accumulation of different proportions of the graph convolution network of each channel,the characteristics of drug molecules can be learned more accurately.Finally,the drug feature vector extracted by the above model is concatenated and fused with the target feature vector extracted by the graph convolution network,and the final predicted value of drug-target affinity is obtained through the full connected layer.On this basis,to further improve the prediction speed of the model and better learn the characteristics of the model,a multi-channel graph convolution drug target affinity prediction model based on multi-head self-attention is proposed in this paper.In this method,the graph data of drug molecules successively pass through the densely connected graph convolution module,the multi-head self-attention module,the multi-channel densely connected graph convolution module,the linear binding layer,the global pooling layer,and the full connection layer to obtain the drug feature vector.Then,the drug feature vector is concatenated with the target feature vector extracted through the graph convolution network,and the final predicted drug target affinity is obtained through the full connected layer.The experimental results show that on the two benchmark data sets,the multi-channel graph convolution model based on multi-heads self-attention is much improved compared with the previous model and gets better prediction results.In conclusion,based on the multi-channel graph convolution network and multi-head self-attention,this paper constructs a prediction model and achieves good prediction performance.The study aims to provide an accurate prediction of the interaction strength between the target and the drug,to decrease the cost and improve the efficiency of the drug development process.