A Novel Model-based Method for Feature Extraction from Protein Sequences for Classification

Representation of amino-acid sequences constitutes the key point in classification of proteins into functional or structural classes. The representation should contain the biologically meaningful information hidden in the primary sequence of the protein. Conserved or similar subsequences are strong indicators of functional and structural similarity. In this study we present a feature mapping that takes into account the models of the subsequences of protein sequences. An expectation-maximization algorithm along with an HMM mixture model is used to cluster and learn the models of subsequences of a given set of proteins.