Prediction Of Protein-RNA Complex Structure And Binding Affinity

Protein-RNA interaction is of great significance in the metabolic process of organisms,and it participates in many biological processes including transcription,translation,and regulation.Currently,more than 240,000 protein-RNA interactions have been measured through high-throughput experimental methods.Compared with known protein-RNA interactions,the number of protein-RNA complex structures in the PDB database is only about 4000.It is time-consuming and labor-intensive to determine the structure of proteinRNA complexes through experimental methods,so it is necessary to develop computational methods to predict the structure of protein-RNA complexes.Traditional protein-RNA complex structure prediction algorithms include free docking and template-based methods for modeling.In protein-protein complex structure prediction,there is evidence that free docking and template-based methods are complementary.Are free docking and template-based predictions of protein-RNA complex structures equally complementary? In the previous template-based protein-RNA complex structure prediction method,RNA structure comparison was done through SARA.However,SARA is dependent on the length of RNA,so it is difficult to judge whether the compared two RNAs are similar.In order to solve this problem,our group developed a size-independent RNA structure alignment algorithm RMalign.In this article,first,in order to verify the reliability of RMalign in more RNA structure,we perform a pairwise comparison of all RNA structures in PDB.After that,we classified all RNAs based on structural similarity and developed an RNA structure classification database RR3 DD.By analyzing the classification of some special RNA structures,we found that RMalign can classify these RNAs well.In addition,the relationship between RNA sequence and structure shows that RNA structure is more conservative than sequence.Therefore,more templates can be found through structural alignment than sequence alignment.Based on this,we replaced SARA in PRIME with RMalign and expanded PRIME2.0,which can only model binary protein-RNA complexes,to PRIME2.1,which can model multiple protein-RNA complexes.When comparing the template-based method and the free docking method,we found that the two methods are also complementary in the structure prediction of protein-RNA complexes.Therefore,we combined free docking and template-based methods to predict the structure of protein-RNA complexes,and developed a protein-RNA complex structure prediction website P3 DOCK.By comparing the current best methods on the three docking benchmarks,the success rates are increased by 3% to 20% for top 10.The success rates on the three different docking calibration data sets are 56%,56% and 61% for top 10,respectively.And then,through statistical analysis of the structure of the protein-RNA complex,we found that modified residues/bases have an impact on the interaction between protein and RNA.The traditional free docking algorithm does not consider modified residues/bases.Therefore,we considered the influence of modified residues/bases on protein-RNA interactions in free docking.In the end,after considering the modification,the near-native structure ranking can be made more advanced.In the prediction of the complex structure with the bound and unbound docking,the success rate with modified residue/base is higher than that of 3d RPC for top 1.Next,according to the characteristics of the convolutional neural network in deep learning,we encoded the structure of interface of the complex into different channels according to the atom type.And then we use the deep learning model to score and sort the decoys.By performing on the integrated docking benchmark for comparison,the success rate of the free state is 3% better than that of 3d RPC for top 10,and can reach 30%.Finally,because predicting protein-RNA binding affinity helps to understand the molecular interaction mechanism of protein and RNA.In order to better fit the binding affinity prediction model,we first updated the protein-RNA binding affinity data set.Based on these data,we trained a model for predicting protein-RNA binding affinity based on structural features.It achieved the Pearson correlation coefficient 0.57 when tested on 41 protein-RNA complexes of PDBbind.