Study On Prediction Of Metal Ion Binding Sites In RNA Structure Based On Deep Learning Methods

Ribonucleic acid（RNA）has important and diverse biological functions in various processes of cell life.Metal ions have been recognized as key factors in controlling RNA folding and are essential for many different functions of RNA molecules.Despite many RNA structures resolved experimentally,accurately detecting them still faces many challenges.Three-dimension convolutional neural networks（3DCNN）can directly learn useful features from original data,which have been applied to the predictions of molecular structure and function.Therefore,based on 3DCNN,we propose theoretical prediction methods for metal ion binding sites in RNA structure at the nucleotide and atomic levels respectively.The main research contents include:（1）A 3DCNN-based method,RNAMG,was proposed to predict the binding sites of Mg²⁺in RNA structure at the nucleotide level,the aim is to preliminarily explore the possibility of using deep learning method to predict metal ion binding sites in RNA.Firstly,a high-resolution structure database of non-redundant RNA-Mg²⁺binding sites（including 113 structures）is constructed by collecting RNA structures from the Protein Data Bank database（PDB）.Subsequently,the local coordinate system is established for each nucleotide,the microenvironment around the nucleotide（atomic distributions of carbon,oxygen,nitrogen,phosphorus and atomic charges）is used as features,and then the tertiary structure of RNA is sent into the 3DCNN model as a 3D image to predict the RNA-Mg²⁺binding site.The accuracy of RNAMG in 5-fold cross validation is0.661.Prediction on an independent test set indicate that in identifying Mg²⁺binding sites in RNA structures,RNAMG has a distinct advantage over the currently proposed advanced nucleotide level-based methods,RBind and RNASite.Matthews correlation coefficient（MCC）of RNAMG is 0.345,which higher than that of RBind and RNASite（MCC is 0.109 and 0.058,respectively）.（2）After the results of using 3DCNN to predict the binding sites of Mg²⁺in RNA structure at nucleotide level was verified,the 3DCNN model named Metal3DRNA was developed to predict the binding sites of different metal ions in RNA structure at the atomic level.Firstly,the non-redundant database of binding sites of different kinds of metal ions(Mg²⁺,K⁺and Na⁺)in RNA structure is constructed from the PDB database.Then,with atom types and charges as input features,Metal3DRNA is trained using the5-fold cross validation on the unbalanced data sets to contract three kinds of models for different metal ions.In independent test set,Metal3DRNA was compared with other atomic-level prediction methods,FEATURE and metalion RNA.In the prediction of metal ion binding sites on 23S r RNA（PDB ID:1HC8）structure by Metal3DRNA model,5 of the 7 experimental Mg²⁺binding sites ranked in the top 7 of the predicted results,slightly better than the other two methods.On other structures in the independent test set,42 of 45 metal ion binding sites were predicted correctly.In addition,a visual analysis of the results was performed on representative cases,and the contribution of features to the predictive decision was explored through significance maps.In conclusion,this study used deep learning to establish new methods for predicting metal ion binding sites in RNA structure at the nucleotide and atomic levels,respectively.Both methods have higher accuracy compared with the current prediction models.This work is of great significance for understanding various key biological processes such as RNA folding,ribozyme-catalyzed reactions and the generation mechanism of related diseases.