New GNN Enhances Protein Representation Learning with Structural Data

Graph-based models are widely employed in protein modeling, yet many existing methods primarily rely on simple sequence adjacency or geometric proximity. These approaches often fail to fully capture the intricate principles governing protein folding, which are heavily influenced by secondary structure elements like alpha-helices and beta-sheets, and stabilizing hydrogen-bond interactions. This research introduces a new secondary-structure-aware graph neural network (GNN) specifically designed for protein representation learning. The model enhances residue-level node representations by augmenting them with secondary structure assignments. Crucially, graph edges are constructed based on hydrogen-bond interactions, which are then filtered by their energetic strength. This innovative design allows the GNN to effectively capture both the local structural context and the critical long-range couplings that are fundamental to protein stability and function. Evaluations on standard protein benchmarks demonstrate consistent improvements over existing graph-based methods, and the resulting graph representations offer superior biological interpretability, aligning well with established structural motifs.