Research on RNA secondary structure predicting via bidirectional recurrent neural network

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• RNA secondary structure prediction is crucial in biological information research.
• Predicting RNA secondary structures with pseudoknots is an NP-hard problem.
• Traditional machine learning methods struggle with variable sequence lengths and sample imbalance in RNA prediction.

Purpose of the Study:

• To address limitations of traditional methods in RNA secondary structure prediction.
• To propose a novel model capable of handling variable sequence lengths and sample imbalance.
• To improve the accuracy and efficiency of predicting RNA secondary structures with pseudoknots.

Main Methods:

• Development of a variable-length dynamic bidirectional Gated Recurrent Unit (VLDB GRU) model.
• Introduction of a flag vector to accommodate sequences of different lengths and utilize surrounding base information.
• Implementation of a weight vector to dynamically adjust the loss function, mitigating sample imbalance.

Main Results:

• The VLDB GRU model demonstrated improved performance compared to existing algorithms on the RNA STRAND dataset.
• Experimental results showed a 4.7% increase in accuracy and an 11.4% improvement in the Matthews correlation coefficient.
• The model effectively utilizes information from both before and after the target base in RNA sequences.

Conclusions:

• The proposed VLDB GRU model effectively handles variable sequence lengths and addresses sample imbalance issues.
• The flag and weight vectors are key innovations enabling improved RNA secondary structure prediction.
• The VLDB GRU algorithm achieves superior detection results compared to other existing algorithms.