Utilizing Domain Knowledge in Neural Network Models for Peptide-Allele Binding Prediction

Abstract

We developed Radial Basis Function Neural Networks (RBFNN) for allele-peptide binding prediction. We explored utilizing prior domain knowledge in order to optimize the prediction. We investigated the effect of encoding of inputs of the RBFNN considering chemical properties of amino acids, detecting motifs in alleles and reducing the dimensionality based on common motifs discovered. We also explored a number of parameters such as the data set size, unknown-binding data generation, model architecture and training algorithms. Our approach improved the prediction accuracy of peptide-allele binding reaching up to 90% for our best models.