Study On Scoring RNA Secondary Structure Based On Deep Learning

RNA plays an important role in the genetic mechanism of living organisms,RNA secondary structure is the primary key to understand RNA function.Due to the lack of complete understanding of RNA,scoring RNA secondary structure is a challenge in bioinformatics.It's difficult to improve the accuracy of scoring functions of long-range RNA sequences in research on RNA secondary structure scoring functions.In this thesis,the bi-directional LSTM(Long Short-term Memory)model is used as the basis of scoring function for shortrange RNA secondary structure,and a recombination layer was designed for long-range RNA sequences to fuse the binding features with the RNA substructure,the characteristics of the depth model are fitted.The training algorithm in this thesis not only uses the genetic algorithm to adjust the proportion of positive and negative target variables,but also uses the k-folder algorithm to cross-verify the model.The recurrent neural network contains more parameters,so that we spend a long time training in the model.Based on this,the Hyper-parameter optimization framework is used to improve the depth sequence model,through the fast parallel search,we obtained a model that is comparable to a manual selection of super-parameters.The Pearson correlation coefficient is about 3% higher than that of the improved bi-directional recurrent neural network,and the violation rate is reduced by about 2%,scoring accuracy has a certain degree of improvement.This thesis not only provides an alternative for scoring RNA secondary structure,but also illustrates how to train a large deep model in an efficient way.