Computer Assisted Peptide Design and Optimization with Topology Preserving Neural Networks

We propose a non-standard neural network called TPNN which offers the direct mapping from a peptide sequence to a property of interest in order to model the quantitative structure activity relation. The peptide sequence serves as a template for the network topology. The building blocks of the network are single cells which correspond one-to-one to the amino acids of the peptide. The network training is based on gradient descent techniques, which rely on the efficient calculation of the gradient by back-propagation. The TPNN together with a GA-based exploration of the combinatorial peptide space is a new method for peptide design and optimization. We demonstrate the feasibility of this method in the drug discovery process.