Visualization And Analysis Of Recurrent Geometric Network Oriented To Prediction Of Protein Structure

Protein three-dimensional structure prediction has always been an important topic in molecular biology,and traditional experimental methods are very complicated and time-consuming.RGN(Recurrent Geometric Networks)as a deep learning model has been successfully applied to the prediction of protein three-dimensional structure.The model uses an amino acid sequence and related PSSM(Position Specific Scoring Matrix)information to predict the three-dimensional structure of the protein backbone corresponding to a sequence and the accuracy of the prediction results can be comparable to the current best method.However,the internal complexity and non-linear structure of the neural networks in RGN make the model itself a "black box",and it is difficult for people to understand the reason for the high accuracy of the network.At present,the explanatory development of neural networks is far behind the development speed of its applications.Researchers urgently need to understand the internal working principle and behavior process of neural networks in order to design better network models.First of all,in response to the "black box problem" of the hidden layer of neural networks,this paper proposes a new method to study the hidden states of neural networks.This method finds the odd and even position characteristics of the hidden states by calculating the similarity of the hidden states at different time steps.It is found that the effect of the backward networks is far greater than that of the forward networks by analyzing the performance of the bidirectional neural networks in RGN at different time steps.Secondly,in response to the problem that the response states of the forward networks is much smaller than that of the backward networks,this paper focuses on the comparative analysis of the hidden layer of the forward and backward networks in the bidirectional neural networks.After comparison,it is found that although RGN uses a bidirectional neural network architecture,the response effect of its forward networks is much smaller than that of the backward networks,and with the time step changes,the response of the forward networks increases suddenly and sharply at the end of the sequence.Through the analysis of the above results,The conclusion given in this paper is that the bidirectional neural network in RGN has the problem of uneven feature learning.Thirdly,in response to the complex and difficult-to-analyze problem of neural network hidden layer,this paper designs and builds a visualization system to analyze the hidden states and other information of RGN.This system improves the efficiency of RGN hidden states analysis and provides convenience for RGN research.Finally,this article is tested on the CASP(Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction)7 datasets,and the experimental results verify the validity of the analysis.