| The prediction of protein structural domains and their boundaries from amino acid sequence data is an open problem of interest to the bioinformatics community. Determining a protein's structural domains experimentally can be very difficult. The assignment of structural domains relies on the time-consuming process of determining the three-dimensional structure. There is a strong potential for computational methods to aid in this and other tasks in structural molecular biology. Computational methods have aided in the prediction of secondary structure and function and the potential exists to predict structural domains from amino acid sequence data. Efforts from other researchers, along with the results presented here indicate that some boundaries can indeed be predicted from the amino acid sequence alone. A new machine-learning based architecture is introduced to predict structural domain boundaries (also referred to as inter-domain linker regions), as defined in a CATH-classified dataset. The key feature of this architecture is a Mixture of Experts (MoE) model. The MoE provides the ability to combine the predictions of individual classifiers, such as artificial neural networks, support vector machines, and naive Bayes classifiers, so as to capitalize on each predictor's strengths and weaknesses. |
