MHC class I molecules are key players in the human immune system. They bind small peptides derived from intracellular proteins and present them on the cell surface for surveillance by the immune system. Prediction of such MHC class I binding peptides is a vital step in the design of peptide-based vaccines and therefore one of the major problems in computational immunology. Thousands of different types of MHC class I molecules exist, each displaying a distinct binding specificity. The lack of sufficient training data for the majority of these molecules hinders the application of Machine Learning to this problem.
We propose two approaches to improve the predictive power of kernel-based Machine Learning methods for MHC class I binding prediction: First, a modification of the Weighted Degree string kernel that allows for the incorporation of amino acid properties. Second, we propose an enhanced Multitask kernel and an optimization procedure to fine-tune the kernel parameters. The combination of both approaches yields improved performance, which we demonstrate on the IEDB benchmark data set.