Protein function prediction is a key task in the field of bioinformatics. Although existing methods based on graph neural networks (GNNs) can utilize protein structure information, traditional graph convolutional networks (GCNs) lack the ability to distinguish residual importance when capturing local structures, making it difficult to accurately identify function related to local associations. To conclude, this article proposes a GCN-GAT fusion module with a learnable weight matrix and a multi head attention mechanism module. Each layer adaptively fuses weighted GCN and GAT to construct a three-layer progressive fusion structure, which retains the ability of GCN to efficiently model local geometric structures and highlight functional key residues through GAT attention. Experiments have shown that the fusion module can effectively enhance the model’s ability to characterize protein structural features, providing better computational tools for protein function prediction.