Study On RNA Secondary Structure Prediction Based On Deep Learning

RNA is the carrier of genetic information.It plays a key role in gene encoding,decoding,regulation,expression and other aspects.RNA secondary structure prediction is the main way to know the biochemical functions.Pseudoknots are the most extensive RNA structure units.RNA secondary structure prediction with pseudoknots is really hard work which has been proved an NP-complete problem.The thermodynamic model is a classical RNA secondary structure prediction method,search for the lowest energy RNA secondary structure.Based on RNA base-pair probability matrix,we propose two improving algorithms with the principle of single-stranded base and base ”triple tie”.Furthermore,this thesis presents a parallel approach to predict RNA secondary structure with pseudoknots.The thermodynamic model can quickly obtain reasonable RNA secondary structures,but it is easy to fall into the local optimum and influenced by the thermodynamic parameters.Deep learning does well in feature mining,can break through the various restrictions in thermodynamic models.Bidirectional recurrent neural network can learn the characteristics of the two aspects of feature information,suitable for deep RNA sequence modeling.This thesis studies modeling RNA sequences by applying deep learning models,such as recurrent neural networks.We propose two deep models for fixed-length RNA secondary structure prediction and variable-length RNA base pair constraint prediction.We combine base pair constraint with parallel ant system and design a new method to predict RNA secondary structure.This approach is compared with several popular RNA secondary structure prediction methods on classic RNA datasets.The experimental results indicate that our method is able to improve the accuracy of RNA secondary structure prediction with pseudoknots by18%.RNA secondary structure has prediction multiple aspects.Selecting the best prediction is an essential step.This thesis proposes a selecting method based on spectral clustering.Experiments indicate that our method is able to pick up better structures.