Accurate Prediction of Protein Secondary Structure By Non-Parametric Models

Proteins are one of the most important parts of an organism because of their vital tasks. Consequently, it is necessary to know both the primary and secondary structure of a protein as closely related to its biological function. Artificial Neural Networks (ANNs) are a useful methodology for secondary structure prediction of proteins. In this study, a generalized regression neural network (GRNN), a probabilistic neural network (PNN) and a backpropagation algorithm (BP) were investigated with different window sizes of amino acid sequences to predict the protein secondary structure from the primary structure. All the outputs of the networks were further combined with another GRNN for possible increase in accuracy rates. The data sets were prepared with alpha-141 and beta-146 hemoglobin chains from the Protein Data Bank. Once the window size was optimized, the overall success rate of GRNN was around 90.2-91.7% for beta chains and 85.9-87.3% for alpha chains. The PNN achieved between 91.2-92% overall accuracy for beta chains and 85.4-86.5% for alpha chains. The BP had an overall accuracy rate of 89.9-92.5% for beta chains, and 86.5-90.3% for alpha chains. Since the networks compared gave similar results, it was concluded that it is very important to choose the optimal window size given a particular network to achieve best accuracy rates.