Improved image based protein representations with application to membrane protein type prediction

With the explosion of protein sequences generated in the postgenomic era, there is a need for the development of computational methods to characterize and classify them as an alternative to the experimental methods that are expensive and time consuming. Although the amino acid chains that constitute proteins are originally symbolic chains they can be converted into numerical sequences and processed as signals. One recent approach represents a protein as a set of images derived from numerical representations of the protein based on the physicochemical properties of amino acids. Then a feature vector is extracted from texture descriptors of the set of images. In this paper we adopt the same approach of representing proteins as sets of images but we propose to generate the images from evolutionary or structural characterization of proteins instead of generating them from physicochemical properties. We also propose the use of an alternative texture descriptor that, in combination with the proposed approach, obtains a significant improvement of classification accuracy in a membrane protein type prediction task.