Protein Model Ouality Assessment Research Based On Graph Neural Networks

Protein is the most important carrier of life activities,and determining the structure of protein is of great significance for understanding its function.In recent years,the use of deep learning methods for protein structure prediction from protein sequences has made great progress.As an essential step in protein structure prediction,protein model quality assessment can not only help to select the best protein conformation from the model pool,but also provide a reference for local optimization of the model,which is the final step in protein structure prediction.In recent researches of protein model quality assessment,the feature representation learning of protein structure using deep learning method is the most effective.When modeling the protein model to be assessed,methods of using sequence representations cannot express spatial proximity between non-adjacent residues in backbone,and meanwhile methods that project structures onto 3D grids and use 3D convolutional networks to learn voxelized feature representations lack rotation invariance and ignore sequential features.Therefore,this work will investigate the use of graph structures to explicitly model the sequence and 3D structure of proteins,and the use of graph neural networks to train feature representations,which are summarized as follows:1.Protein models are represented as linear graphs,in which the nodes represent residues,and the edges represent the covalent bonds that constitute the protein backbone,connecting consecutive residues on the primary structure.In addition,in order to describe the spatial relationship between non-adjacent residues,an edge is formed for two residues within a chemically reasonable distance,such an edge represents contact between the two residues.Model proteins as graphs of residues,bonds,and contacts,and add secondary structure and coevolution information to the graph.2.In order to learn the local and global feature representation of proteins,this research will build a neural network composed of Graph Transformer feature extractor and message passing layer,and improve Graph Transformer by adding edge features to the computation of attention score,which is more suitable for the protein graph of this study.In the message passing layer,the global feature representation is added to the training,rather than obtained by pooling local features,thereby optimizing the learning of global features.3.Use a more refined graph structure: Model a protein conformation into a graph with atoms as nodes,and edges represent covalent bonds and contacts between atoms.A pooling layer is introduced to pool atomic-level feature representations into residue-level feature representations,and generate a new coarse-grained graph structure,which is then input into the model in this paper for training.The experimental results of this paper show that,based on protein graphs,the graph neural network integrating Graph Transformer and graph convolutional layers can effectively extract local features,characterize the correlation between adjacent nodes in the local neighborhood and their contributions to the local environment to different degrees,and effectively spread local information to larger neighborhoods.