Fragment-based backbone sampling & neural net derived potentials to design protein-binding peptides 

Presented by Sebastian Swanson on June 14th 2023
Abstract:
Computational methods for de novo peptide design could unlock the ability to target previously difficult to bind sites on natural proteins, but the large space of backbone structures and inaccuracies in energy functions pose major challenges. We present a framework for de novo peptide design that combines fragment-based backbone sampling with graph neural network-derived scoring potentials. We generate candidate peptide backbone structures, which we refer to as seeds, using tertiary fragments from known protein structures. We score peptide seeds with COORDinator, a graph neural network for fixed-backbone sequence design, allowing us to identify hot-spot peptide seeds. We demonstrate through computational benchmarks that hot-spot seeds are predicted to form interactions that are as good or better than native peptides to a diverse set of binding sites. We also present a case study, where we use our method to design peptide binders of a bacterial toxin, some of which are predicted to bind with high confidence by AlphaFold. Altogether, this work demonstrates the promise of our approach and highlights some of the unique challenges that remain in order to develop more robust methods for designing protein-binding peptides.