Currently, accurately identifying autism spectrum disorder (ASD) remains a significant challenge. Most existing methods primarily focus on handling single atlas brain data for ASD identification, without thoroughly investigating the significance of utilizing multi-atlas data in ASD identification. Hence, we propose a learning framework called MM-GNN that combines multi-atlas and graph neural networks to enhance the identification of ASD. Firstly, to effectively fuse multi-atlas features, we design a multi-atlas feature representation module based on hyper-graph edge representation learning, which maps different brain atlas features into a unified feature representation space. Secondly, we design channel and spatial attention mechanisms to adaptive learn the weights of different atlases for multi-atlas feature fusion. Lastly, we design the graph neural network model to learn the complex relationship between different subjects of multi-center data. Through 10-fold cross-validation, we evaluate the proposed framework and show that it surpasses state-of-the-art performance on the ABIDE dataset, with an average accuracy of 91.96% and an AUC-score of 97.22%. Our results demonstrate that the proposed method applies more to multi-center and multi-atlas autism identification.