An Architecture Combining Bayesian segmentation and Neural Network Ensembles for Protein Secondary Structure Prediction

Abstract

A combined architecture of Bayesian segmentation along with ensembles of two layered feedforward network has been built and tested on widely studied two non membrane, non homologous databases comprising of 480 and 608 protein sequences respectively. In the first stage, Bayesian segmentation is used to infer sequence/structure relationship in terms of structural segments which is well suited to model non-local interactions among segments. The probability scores for the three structural states (helix, sheet and coil) of each residue obtained from the Bayesian segmentation has been used as the inputs at the second stage to a feedforward neural network. The neural network is trained with the sliding window comprising of the scores of seven consecutive residues along with additional inputs for physicochemical properties of the residues where the prediction is made for the central residue. The key aspect of the model is inclusion of physicochemical properties of the amino acids at the second stage. An ensemble of neural networks have been trained in second stage based on the posterior probabilities approach to determine the number of neural networks. This model achieves a Q3 accuracy of above 71% which is one of the highest accuracy values for single sequence prediction methods.