Learning Protein Structure Representation with Orientation-Aware Networks

The intricate three-dimensional structures of proteins dictate their diverse biological functions. Accurately representing these structures, particularly the fine-grained orientational relationships between amino acids, is crucial for understanding protein mechanisms and practical computational protein analysis. Current methodologies often fall short in capturing these essential geometric details. We introduce Orientation-Aware Graph Neural Networks (OA-GNNs), a novel deep learning framework that explicitly models local and global geometric characteristics, including inner-residue torsion angles and inter-residue orientations. OA-GNNs achieve this by uniquely extending neural network weights from scalars to 3D directed weights and by implementing an equivariant message passing paradigm that ensures SO(3)-equivariance. This approach allows for a richer, more geometrically meaningful processing of protein structural data. Comprehensive experiments demonstrate that OA-GNNs significantly outperform existing methods in sensing orientational features and achieve state-of-the-art performance across diverse computational biology tasks, including residue identification, computational protein design, model quality assessment, and protein function classification. Our findings highlight the power of orientation-aware learning and establish OA-GNNs as a versatile and robust tool for advancing our understanding of protein structure–function relationships and for developing new therapeutic and biotechnological solutions. The code is available at https://github.com/Ced3-han/OAGNN/tree/main .