Graph Neural Networks (GNNs), with their remarkable capabilities in learning graph features, have achieved success in graph classification tasks. However, the Message Passing GNNs (MP-GNNs), sharing the feature aggregation mechanism with the 1-Weisfeiler-Leman (1-WL) graph isomorphism test, suffer from limited expressiveness in downstream tasks, such as graph classification. To enhance the expressive power of GNNs, we propose a substructure-aware and high expressive graph neural network model. Firstly, Ego substructures are extracted from a graph and fed into a MP-GNN to obtain substructure-level features. Then, a count-sensitive neighborhood hashing method is introduced to encode the structural characteristics of Ego substructures, which are then injected into the corresponding substructure-level features to enhance the model’s expressiveness. Next, an Ego-substructure-aware Transformer is utilized to extract global features from the graph, which use a <CLS> token to fuse the interaction information among graphs’ nodes, thereby generating graph-level feature vector. Finally, we apply a Multi-Layer Perceptron (MLP) followed by a softmax layer to classify graphs. Theoretical analysis demonstrates that the proposed SHGNN is more powerful than 1-WL test. Experimental results show that, compared to the state-of-the-art graph deep learning methods, our model not only outperforms 1-WL in expressiveness, but also achieves higher classification accuracy.</CLS>