Refined Protein Contact Map Prediction Based On Graph Neural Network

Protein is an important component of organism tissues,and cells as well.It is a macromolecular substance formed by one or more amino acid sequences through spiraling.,folding and coiling.It participates in almost every physiological function of life.The three-dimensional structure of a protein is the basis for its various functions.Therefore,the study of protein structure helps to understand how protein function,to reveal the mechanisms by which proteins interact with each other or with other molecules,and promotes the development of biology and medicine.The protein residue contact map prediction plays a key role in protein structure study.The contact map contains the topological information of the three-dimensional structure of the protein,provides necessary constraints of the protein structure,helps to obtain a more accurate three-dimensional protein conformation,and therefore is widely used in protein structure evaluation and prediction,protein structure matching,and other tasks.There are three main types of methods for protein residue contact map prediction,namely direct coupling analysis methods,machine learning-based methods,and deep learning-based methods.The disadvantage of the direct coupling analysis method is that the prediction accuracy is not very high on condition that the related homologous proteins is not sufficient.Although machine learning models can solve the above problems,they cannot perform large-scale model training with large amounts of data.Most of the deep learning models for contact prediction between residues are based on the convolutional neural network(CNN)architecture,the convolution kernel structure of which is easier to capture local information.However,the protein residue network is a complex non-Euclidean structure,and the convolutional neural network cannot capture the characteristic information of residues completely,so its accuracy has not yet reached the level of completely replacing biological experiments.Graph neural network(GNN)is a novel deep learning method that directly acts on graph structure,and has good potential to be applied to protein contact prediction methods.This paper proposes a Refined Contact Map Prediction Model(RCMPM),which uses existing methods to obtain a rough contact map between protein residues,and then combines graph convolutional network and residual network structure to refine the relationship between residues of the contact map.Among them,the rough contact graph is used to construct a protein graph consisting of amino acid nodes,while graph convolutional network is used to better capture global information,that is,the longrange correlation of protein sequences.The residual network structure in the model contains a one-dimensional residual neural network module and a two-dimensional residual neural network module,which are more suitable for efficient processing of local information.Three sequence features,namely position-specific scoring matrix,secondary structure and solvent accessibly are selected as the input of graph neural network node and one-dimensional residual neural network module,and the interresidue scores,the direct evolution coupling information,the mutual information,and the contact potential between residues are selected as the input of the 2D residual neural network module.In the experiments,the model is trained using the training set in the PDB25 dataset,and the model is validated using the testing dataset in PDB25 dataset and CASP10,CASP11,and CASP12 datasets.Experimental results show that our proposed RCMPM method significantly improves the initial rough contact map,especially on the prediction results at the long-range level.Compared with existing methods,the RCMPM method has the best overall performance on the four datasets and is also the most stable one.It shows that the proposed RCMPM method is effective in the protein contact refinement prediction task.