Predicting RNA Secondary Structure via Adaptive Deep Recurrent Neural Networks with Energy-based Filter'

Weizhong Lu ¹, Weizhong Lu ¹, Ye Tang ¹, Hongjie Wu ^1,2,*, Hongmei Huang ¹, Jing Qiu ¹, Haiou Li¹,

1 School of Electronic and Information Engineering, Suzhou University of Science and Technology, Suzhou 215000, China 2 2Anhui Key Laboratory of Intelligent Building Energy Efficiency, Anhui Jianzhu Q3 University, Hefei 230601, China *Correspondence should be addressed to Hongjie Wu; Hongjie.wu@qq.com

1.Model framework

Figure 1. Adaptive LSTM model framework (click on the link for HD image)

2.Datasets

a)ct file

3.Source code

4.Experiment result

a) Comparison between Adaptive and Fixed LSTM.

ii.output_adaptive

b) Comparison between adaptive-LSTM with and without energy-based filter

result

Fig. 2 Scatter of accuracy comparison between adaptive-LSTM with and without filter (click on the link for HD image)

c) Comparison between adaptive LSTM and other three classical methods

i.Adaptive LSTM with energy-based filter

iv.centroidfold

d) Case study on a sequence with pseudoknots

result

Fig. 3 Native secondary structure of RFA_00633 (click on the link for HD image)

Fig. 4 Predicted secondary structure of RFA_00633. a ProbKnot. b Cylofold. c Centroidfold. d Adaptive LSTM with energy-based filter (click on the link for HD image)