Hybrid Model Of Neural Network And Hidden Markov Model For Protein Secondary Structure Prediction

Ou-Yan Shi,Hui-Yun Yang,Jing Yang,Xin Tian
2007-01-01
Abstract:Protein secondary structure prediction is an important step towards understanding how proteins fold in three dimensions. Using multiple sequence alignments, two layers NN - based method gets higher prediction accuracy. But window - based approach in NN - based method has the disadvantage of only considering the local information. Recent analysis by information theory indicates that the correlation between neighboring secondary structures are much stronger than that of neighboring amino acids. So we use a 7 - state HMM to replace the second layer network. 496 proteins selected from the dataset CB513 are used in a 7 - fold cross validation. The hybrid model appears to be very efficient, with Q(3) score of 75.96% and SOV of 71.27%, more than 0.96% and 0.45% above two layers NN based method. This hybrid model not only captures the local information, but considers the long - distance information. So it can get higher prediction accuracy.
What problem does this paper attempt to address?