Graph Neural Networks for Protein Structure Prediction via Fragment Assembly

Abstract

Predicting protein three-dimensional structures from their amino acid sequences remains a grand challenge in computational biology. Traditional methods have struggled to accurately capture the complex, long-range interactions that govern protein folding. This work proposes a novel approach utilizing Graph Neural Networks (GNNs) to address this challenge through a fragment assembly paradigm. We hypothesize that proteins can be effectively predicted by learning to assemble smaller, interacting fragments based on their local structural characteristics. Our GNN learns to represent individual protein fragments as graphs, capturing their local interactions via node features (amino acid types, residue connections) and edge features (distances, angles). The network then predicts the optimal assembly order of these fragments, ultimately generating a predicted protein structure. This approach avoids the need for explicit conformational search and leverages the powerful representation learning capabilities of GNNs. We demonstrate the feasibility and potential of this approach, outlining a framework for future development and exploration.