Protein-protein interaction sites are specific surface regions that mediate contacts with partner proteins and are critical for understanding cellular mechanisms and guiding drug discovery. In recent years, graph neural networks (GNNs) and Transformers have become promising tools for protein-protein binding sites prediction. However, most existing methods primarily emphasize semantic representation learning of residues, while their topological organization (e.g., adjacency matrix, positional encoding) may still remain too rigid to effectively adapt to intrinsic conformational changes involved in protein binding. To address this, we introduce COAT-GNN, a GNN-Transformer model with CO-operative Attribute learning and Topological optimization for binding sites prediction. COAT-GNN introduces a physics-inspired and geometrically interpretable attention mechanism that models residue-residue interactions as driving forces in a cooperative learning process: on the one hand, residue features are used to estimate pairwise interactions, which in turn guide their coordinate updates by “pulling” residues toward energetically favorable positions (Attribute \(\rightarrow \) Topology); on the other hand, each residue’s features are refined through localized message passing based on its dynamically updated neighbors (Topology \(\rightarrow \) Attribute), thus accommodating the next round of evolution. This dynamic learning process is embedded in an end-to-end framework reliably guided through extrinsic supervised learning signals, thus effectively steering the self-organizing conformational search in the residue interaction space. Extensive results and ablation studies demonstrate the promising performance and robustness of COAT-GNN.