Abstract:
Many proteins are composed of multiple structural domains. These domains can have important structural or functional properties. When a protein's sequence, but not struct...Show MoreMetadata
Abstract:
Many proteins are composed of multiple structural domains. These domains can have important structural or functional properties. When a protein's sequence, but not structure, is known, being able to predict the division of the sequence into its domains may ease structural or functional analysis of the protein. However, predicting domain boundaries from sequence is still an open problem. This research uses machine learning approaches to the prediction problem. It starts with several standard methods -naive Bayes, support vector machines, and artificial neural networks. The results of these methods are combined using a mixture of experts (MoE) approach that gives the final boundary predictions. We show both a simple MoE approach, and one using context in the form of a window of predictions from the machine learning methods. Both MoE methods, especially the windowed MoE, show greatly increased predictive performance relative to the individual machine learning predictors.
Date of Conference: 14-17 October 2007
Date Added to IEEE Xplore: 05 November 2007
ISBN Information: