Artificial intelligence systems based on texture descriptors for vaccine development


The aim of this work is to analyze and to compare several feature extraction methods for peptide classification that are based on the calculation of texture descriptors starting from a matrix representation of the peptide. This texture-based representation of the peptide is then used to train a support vector machine classifier. Experimental results using our new descriptors dramatically outperforms previous results that have been reported using texture descriptors for peptide representation. In our experiments, the best results are obtained using local binary patterns variants and the discrete cosine transform with selected coefficients. In addition, we perform experiments that combine standard approaches based on amino-acid sequence. The experimental section reports several tests performed on a vaccine dataset for the prediction of peptides that bind human leukocyte antigens on a human immunodeficiency virus (HIV-1). Finally, to better evaluate the existing approaches in a realistic scenario, we perform an extensive evaluation for HIV-1 protease cleavage site prediction using two large datasets, HIV-1 PR 3261 and HIV-1 PR 1625, according to a blind testing protocol where one dataset is used for training and configuring parameters and the other for testing system performance. Experimental results confirm the usefulness of our novel descriptors.

Keywords peptide classification; vaccine development; HIV-1 protease prediction; locally binary patterns; discrete cosine transform; support vector machine.

[full paper]