Research On Prediction Methods Of Drug-related Protein Interactions For Multi-layer Heterographs

The development of new drugs is a time-consuming and expensive process,people are looking for new uses for approved drugs to reduce development time and save research costs.Predicting drug-protein interactions（DPIs）is an important step in drug development and repositioning.Applying computational methods to predict DPIs can narrow the scope of screening and provide biologists with high-probability interaction candidates.Therefore,how to accurately and effectively predict DPI is a problem worthy of research and exploration in the field of bioinformatics.In this paper,three computational models are proposed to predict DTIs based on the multi-layer complex graph structure composed of drug and protein related data.The main tasks completed includes:（1）Aiming at the research on the multiple connections in heterogeneous graph,this paper proposes a drug-protein interaction prediction model DNDPI based on nonnegative matrix factorization and dilated convolution.First,we construct a drug-proteindisease heterogeneous graph to connect multiple similarities,interactions,and associations between nodes.Then,we define the multi-scale attribute representation of drug-protein node pair and the global topological representation of the node.The multiscale attribute representation of node pair is obtained by applying dilated convolutions of different receptive field sizes to the original information embedding,which integrates different inter relationships and intra relationships between drugs and proteins,and learns multi-scale features.The global topological representation is obtained based on nonnegative matrix factorization and relation-level attention mechanism,which captures complex topologies in heterogeneous graphs and adaptively adjusts the contribution of different topologies.Finally,the final interaction prediction score of drug-protein pair is obtained by weighted integration of the interaction scores obtained from the two representations.The results of multiple evaluation metrics and case studies demonstrate that DNDPI outperforms six other state-of-the-art drug-protein interaction prediction models.（2）Aiming at the research on multi-modality similarities of drugs and proteins and multi-order neighbor topological information,we propose a drug-protein interaction prediction model ALDPI based on convolutional neural networks and graph convolutional autoencoders.First,we calculated the multi-modality similarities of drugs and proteins,which reflect the degree of similarity between two drugs（two proteins）from different perspectives,and exploited the interactions between drugs and proteins and the multi-modality similarities to construct drug-protein heterogeneous graph.Then,we propose an adaptive topology graph learning module,which is able to transform heterogeneous graphs into multiple new graph structures formed by different neighbor topologies.We apply independent graph convolutional autoencoders to learn node representations based on different-order neighbor topologies,and assign higher weights to topology representations containing more information through an attention mechanism at the topological representation level.The attribute representation of drug-protein node pair obtained by a multi-layer convolutional neural network integrates multi-modality similarity information and interaction information.Finally,a fully connected neural network is used as a classifier,and the learned topological representations and attribute representations are used to predict drug-protein interaction scores.The comprehensive experimental results show that ALDPI has superior performance than other prediction models.（3）Aiming at the research on the different levels data characteristics related to drugs and proteins,we propose a multi-level autoencoders based drug-protein interaction prediction model MAEDPI.First,we constructed a heterogeneous graph using the interaction and association connections between drugs,proteins and diseases.Then,we encode and embed from the drug-protein heterogeneous subgraph level,the pairwise attribute level,and the topological connection level.Due to the different emphases and characteristics of the three levels,we define three different encoding strategies to learn deep representations of three levels respectively.Finally,the learned three-level representations are fed into a fully connected neural network to predict drug-protein interaction scores.Experimental results show that both AUC and AUPR of MAEDPI are superior to other prediction models,and the case study results demonstrate its ability to discover potential drug-protein interaction candidates.