Computational Prediction and Mapping of Protein Folding Pathways

Abstract

Protein structures are dictated by their sequences, but the mechanisms underlying folding remain ambiguous. Various computational approaches exist to investigate protein folding, but they are often “black-box” tools that only predict native structures. Here, we introduce StepFold, a tool to rapidly explore the fold space by traversing contact maps. By representing 3D structures on a 2D grid, contact maps offer dimensional simplicity through which probabilistic calculations can be performed upon structures. StepFold integrates empirical statistics from experimentally derived structures to predict folding as a series of residue interactions influenced by their local contexts. By incorporating the blob-based model, StepFold generates grounded folding pathways and gives insight into how contacts beget complex folds. The results of this paper show that StepFold can rapidly and efficiently recreate native contact maps through blob-based folding. While its capacity for de-novo structure prediction is limited, StepFold can reproduce structures with an accuracy of over 91% for predicted contacts, while capturing over 62% of those in the native structure. StepFold is both rapid and scalable to large sequences, with a mean runtime of approximately 173 seconds per 1000 folding steps under default conditions. While improvements to the underlying probabilistic model are needed to improve prediction performance, StepFold can already give insights into how local folds cumulatively create complex tertiary structures.