Comparison of feed-forward and recurrent neural networks in predicting thermostable protein temperature

Engineering thermostable enzymes have received attractable research interests and understanding the features involved in thermal stability is very important and here we compared two neural networks modelling (feed-forward and recurrent) on 2938 enzymes of meso- and thermophilic proteins to determine the features involved in this process. We randomly divided records into 10 parts each consist of 294 records. The process repeated 10 times and the accuracy for each repeat and total accuracy were calculated. To investigate the effects of the feature selection on the neural networksbehavior, all models were also run with feature selection (stepwise regression) criteria. The results showed in feed-forward neural network, the best overall accuracy (0.90) obtained when the hidden layer was 3 and the number of neurons in input layers was 50, 20 and 10. of hidden layers was 3 while its accuracy in predicting true, false and overall records were 0.824, 0.935 and 0.892, respectively. In recurrent (Elman) neural networking, the figures were 0.809, 0.946 and 0.894, respectively. The results showed there is no significant difference (p > 0.95) between feed-forward and Elman neural networks and no difference found when stepwise regression feature selection used. When the best networks run on another dataset of 60 protein with known temperatures, the best accuracy in determining the right temperature (97.83%) gained in Elman network and the worst one in recurrent with 2 hidden layers. The findings confirmed both networks are suitable for predicting thermostable proteins and feature selection decreases the burden of system.

Keywords:

Bioinformatics , modeling , protein , thermostability