Fully Differentiable Full-Atom Protein Backbone Generation

Abstract

The fast generation and refinement of protein backbones would constitute a major advancement to current methodology for the design and development of de novo proteins. In this study, we train Generative Adversarial Networks (GANs) to generate fixed-length full-atom protein backbones, with the goal of sampling from the distribution of realistic 3-D backbone fragments. We represent protein structures by pairwise distances between all backbone atoms, and present a method for directly recovering and refining the corresponding backbone coordinates in a differentiable manner. We show that interpolations in the latent space of the generator correspond to smooth deformations of the output backbones, and that test set structures not seen by the generator during training exist in its image. Finally, we perform sequence design, relaxation, and ab initio folding of a subset of generated structures, and show that in some cases we can recover the generated folds after forward-folding. Together, these results suggest a mechanism for fast protein structure refinement and folding using external energy functions.