Abstract
Protein sidechain conformation prediction, or packing, is a key step in many in silico protein modeling and design tasks. Popular protein packing methods typically rely on approximated energy functions and complex algorithms to search dense rotamer libraries. Inspired by the recent success of deep learning in protein modeling tasks, we present ZymePackNet, a graph neural network based protein packing tool that does not require a rotamer library, scoring functions or a search algorithm. We train regression models using protein crystal structures represented as graphs, which are employed sequentially to “germinate” the sidechain starting from atoms anchoring the protein backbone to the sidechains’ termini, followed by an iterative refinement stage. ZymePackNet is fast and accurate compared to state-of-the-art protein packing methods. We validate our model on three native backbone datasets achieving a mean average error of 16.6°, 24.1°, 42.1°, and 53.0° for sidechain dihedral angles (χ1 to χ4). ZymePackNet captures complex physical interactions such as π stacking without explicitly accounting for it in the model; such effects are currently lacking in the energy terms used in traditional packing tools.
Contact abmukho{at}vt.edu
Supplementary information Supplementary data are available at Bioinformatics online.
Competing Interest Statement
The authors are listed as inventors on a patent application related to this subject matter, PCT/CA2022/051612, filed by Zymeworks Inc.