Predicting Fold-Switching Protein Behavior Using Graph Neural Networks

Recent studies have revealed an increasing number of fold-switching proteins, which defy the conventional belief of “one sequence, one structure.” Fold-switching proteins are vital in cellular signaling and regulatory pathways, transitioning between distinct structural conformations in response to environmental cues. It is estimated that up to 4% of Protein Data Bank proteins switch folds, but many more likely remain undiscovered. Despite their biological importance, these proteins remain understudied and can shed light on the dynamics of protein folding, potentially improving protein structure predictions. There is an acute need for prediction methods that can identify fold-switching behavior to expedite discovery and characterization of these proteins. Previous research indicates that it is essential to incorporate complex relationships within protein sequences and structures. This study introduces an approach using Graph Neural Networks (GNNs) to capture intricate relationships among amino acids, including spatial dependencies and structural motifs, enabling more accurate predictions of fold-switching behavior. The methodology involves using physicochemical and sequence-based information, which serve as features for Graph Convolutional Network layers. The GNN model identified fold-switching proteins with an accuracy of 75%, marking a significant improvement over previous methods by utilizing a generic machine learning-based classifier that demonstrates superior performance and broad applicability in predicting fold-switching behavior across diverse protein sequences. This work has practical implications for drug discovery, protein engineering, and synthetic biology, and paves the way for future research by enabling the development of more sophisticated models to unravel the complexities of protein folding and function, thereby enhancing our understanding of cellular processes and disease mechanisms.