Neural network approach to predict RNA secondary structures

Ribonucleic acid (RNA) secondary structure predictions based on stochastic context-free grammar (SCFG) models are very complex. This paper presents a BP neural network approach for predicting RNA secondary structures based on a new representation of the RNA structure information. The new format for the secondary structure prediction results can be easily converted to the commonly-used CT format. Test results obtained with tRNA training and testing datasets show that the approach has higher prediction accuracy and greater correlation coefficients than the two best-performance SCFG models. Since computational complexity for heuristic neural network approaches are relatively simple, the method can be used to solve secondary structure prediction problems of long RNA sequences with lengths greater than 1000 nt, which are difficult with traditional folding algorithms.