SE(3)-Equivariant Energy-based Models for End-to-End Protein Folding

Accurate prediction of protein structures is critical for understanding the biological function of proteins. Nevertheless, most structure optimization methods are built upon pre-defined statistical energy functions, which may be sub-optimal in formulating the conformation space. In this paper, we propose an end-to-end approach for protein structure optimization, powered by SE(3)-equivariant energy-based models. The conformation space is characterized by a SE(3)-equivariant graph neural network, with substantial modifications to embed the protein-specific domain knowledge. Furthermore, we introduce continuously-annealed Langevin dynamics as a novel sampling algorithm, and demonstrate that such process converges to native protein structures with theoretical guarantees. Extensive experiments indicate that SE(3)-Fold achieves comparable structure optimization accuracy, compared against state-of-the-art baselines, with over 1-2 orders of magnitude speed-up.