Fast fixed-backbone protein sequence and rotamer design

Namrata Anand, Tudor Achim

Keywords: protein design, imitation learning, attention, graph networks

Abstract: Protein fixed-backbone sequence design is an important task in computational protein design, and being able to quickly and accurately modify or redesign side-chains is a useful subroutine in the context of functional design problems such as ligand binding site, enzyme, and binder design. We present a fast and accurate learned method for protein fixed-backbone sequence and rotamer design. We find that a graph attention model for joint rotamer and sequence prediction trained on-policy via imitation learning can produce a distributions of accurate sequences for target backbones. We show that this method generalizes to design sequences onto novel generated backbones.

One-sentence Summary: We report an algorithm for fast fixed-backbone protein sequence design.