| Protein secondary structure prediction methods aim to accurately predict the structure of a protein given knowledge only of its primary sequence. In this thesis, we investigate a new approach to the prediction of the protein secondary structure which creates species-specific predictors instead of using a single structure predictor trained using data pooled from multiple species. The underlying hypothesis that protein folding is influenced by species-specific differences is first investigated through a comparison of protein chain sequence and structure composition for 12 species representing all six Kingdoms of life. Next, various neural networks are trained with species-specific data to determine if there exists a particular neural network architecture that yields optimum prediction accuracy for a particular species. Through evaluation of five different network architectures, results show that the performance of Elman networks surpass other network architectures for most of the species. Elman networks are then trained with species-specific sequence and structure data. Five-fold cross-validation results over 12 species reveal that species-specific predictors are more effective than predictors trained on protein data pooled from multiple species. Interestingly, when an exact match between the test and train species is not available, results over 16 new species indicate that there is preference for predictors trained on phylogenetically related species. Lastly, we show that voting among several species-specific classifiers provides the highest classification accuracy. To my knowledge, this work represents the first investigation of species-specific neural network protein secondary structure prediction systems. |
