GDFold2: a fast and parallelizable protein folding environment with freely defined objective functions

Abstract

An important step of mainstream protein structure prediction is to model the 3D protein structure based on the predicted 2D inter-residue geometric information. This folding step has been integrated into a unified neural network to allow end-to-end training in state-of-the-art methods like AlphaFold2, but is separately implemented using the Rosetta folding environment in some traditional methods like trRosetta. Despite the inferiority in prediction accuracy, the traditional approach allows sampling various protein conformations compatible with the predicted geometric constraints, partially capturing the dynamical information. Here, we propose GDFold2, a novel protein folding environment, to address the limitations of Rosetta. On the one hand, GDFold2 is highly computationally efficient, capable of accomplishing multiple folding processes in parallel within the time scale of minutes for generic proteins. On the other hand, GDFold2 supports freely defined objective functions in order to fulfill diversified optimization requirement. Moreover, we propose a quality assessment (QA) model to provide reliable prediction on the quality of protein structures folded by GDFold2, thus substantially simplifying the selection of structural models.