Transmembrane Protein-Binding Domain Prediction Based On Graph Neural Networks

Transmembrane proteins are proteins that exist on the cell membrane and regulate the exchange and communication of substances inside and outside the cell.Signaling molecules inside and outside the cell can control various physiological activities of the cell by binding to transmembrane proteins for information transfer and material transport.Therefore,studying transmembrane protein-binding provides insights into the underlying biochemical mechanisms in organisms,reveals molecular mechanisms for signaling and regulation within cells,and provides fundamental research for the development of new drugs and treatments for diseases.However,due to the paucity of data and structural specificity of transmembrane proteins,large-scale biological experimental techniques for transmembrane binding are currently scarce,so there is an urgent need to develop efficient and reliable computational tools to support transmembrane binding.In this paper,we aim to solve the problem of insufficient data of transmembrane protein structure and binding site samples,predict transmembrane protein binding site by using protein sequence characteristics,understand the relationship between transmembrane protein binding and sequence characteristics and structure,and obtain new transmembrane binding sample representation.Based on these results,we explored binding regions with structural stability and specificity on the local surface of transmembrane proteins based on basic knowledge of protein structure-determining functions.The binding domain consists of a set of spatially similar residues on the protein surface and contains binding sites related to binding action,which can significantly characterize the structural and biochemical properties of protein binding action.Therefore,in this paper,we first build a deep learning prediction model for transmembrane protein binding sites to verify the effectiveness of deep learning for binding feature extraction.Based on this,a binding domain prediction model is designed to conduct more in-depth binding studies.For the prediction of transmembrane protein binding sites,we designed a prediction model of transmembrane protein binding sites based on self-attention mechanism and graph neural network.A transmembrane protein-binding locus dataset was established to construct a complete sequence residue map of transmembrane protein based on transmembrane protein sequence.Instead of the traditional sliding window-based method to extract the local functional features of binding sites,the crossmembrane protein embedding representation is obtained by using the self-attention mechanism to extract the contextual relationship characteristics between residues.Finally,a prediction model of transmembrane protein binding sites was constructed by using the Graph Attention Network.The final validation using the method on the benchmark dataset resulted in a prediction accuracy of0.8491.an AUC value of 0.6216;a Matthew correlation coefficient of 0.3498;a recall of 0.5891;and an F-score of 0.6207.For transmembrane protein binding domain prediction,a binding domain prediction model based on attention-pool mechanism and residual graph convolution neural network is designed in this paper.The attention coefficient matrix between nodes is extracted by using the combined site prediction model,and the potential edge relationships in the whole sequence residual matrix of transmembrane proteins are found by integrating the location neighborhood relationships.A convolution neural network(CNN)model with residual operation is designed to extract the node importance fraction in the subgraph.In this way,Top-K pooling is carried out to obtain the important characteristic information of the key functional region of binding and to predict the binding domain.The final validation using the method on the benchmark dataset resulted in a prediction accuracy of 0.8191.an AUC value of 0.8657;a Matthew correlation coefficient of 0.5737;a recall of 0.5263;and an F-score of 0.8651.In conclusion,this paper aims at obtaining the characteristic and locus information of transmembrane protein and constructing the whole sequence residue map of transmembrane protein in order to solve the problems of difficulty in obtaining transmembrane protein structure data and the lack of local functional region characterization in traditional methods.This paper extends from the traditional research of binding sites to the exploration of binding local areas,not only relieves the existing problems in the research of transmembrane protein binding,but also breaks the bottleneck of the traditional research thinking.Extensive prediction and systematic study of membrane protein binding will help to explore more functional and biological mechanisms of transmembrane proteins,further advance the discovery of protein drug targets and drug development,and provide new ideas and directions for transmembrane protein research.