Comparative Visual Analysis Of Protein Prediction Results For Recurrent Geometric Network

The spatial structure of a protein determines the function of the protein,which is essential for inferring the evolutionary relationship between protein structures,drug discovery and protein design.The advancement of machine learning has promoted the development of protein structure prediction,increasing the prediction speed from months,days,and hours to units of seconds and milliseconds,and the number of codes has also been reduced from a million lines to a few thousand lines.However,whether the prediction accuracy and stability of the protein structure meet people's needs requires further analysis.This paper takes the protein tertiary structure data predicted by the Recurrent Geometric Network(RGN)as the main research object,and conducts research from the perspective of structural comparison,visual analysis,and whether the prediction results of the RGN network meet the needs of experts in the field.Firstly,aiming at the problem that the limitations of RGN network output results in visual analysis and conformational analysis,a multi-angle analysis method for RGN network output results is proposed in this paper.This method provides standard data interfaces and data support for protein visualization tools and the visual analysis framework of this paper from the above two perspectives,including the conversion of protein structure standard files,and the main influencing factor of protein structure conformation-torsion angle data conversion.In addition,this method not only preserves the basic information of protein structure data,but also facilitates the analysis of the visual framework.Secondly,in view of the question of whether the RGN network results meet the needs of domain experts,this paper proposes a multilevel comparative visual analysis method for RGN network prediction results.This method mainly analyzes the level of prediction accuracy,the level of structural difference and the level of structural stability of the RGN network,including the analysis of the protein structure similarity index for its prediction accuracy;the analysis of its structural difference from the distance deviation and the change of torsion angle;the analysis of whether its conformation conforms to the biological level change from the aspect of conformational stability.Finally,this paper designs a multilevel visual analysis framework based on the methods proposed above.Through a variety of visualization methods,the reasons for the differences in the prediction structure of the RGN network are analyzed,as well as the future improvement direction of the RGN network,which verifies the validity of the experimental method.