Abstract <p>Functional connectivity analysis of brain networks has provided valuable insights for brain disorders diagnosis. Recent studies have focused on collaborative learning with multiple brain atlases to overcome the limitations of single-atlas information. However, these approaches often overlook sufficient interaction and consistency among multiple atlases, as well as information redundancy resulting from multi-atlas fusion. We propose a unified multi-atlas learning framework (BrainUMA) with hyper-connectivity network learning for brain disorders diagnosis, which consists of two key stages: hyper-connectivity network construction, and cross-atlas HCN interactions. We employ FCN for hyper-connectivity network construction and propose a novel hyper-connectivity network construction strategy, which includes both the hypergraph structure construction and node feature learning. Meanwhile, to sufficiently model interactions across multiple atlases, we propose a feature disentanglement method that disentangles disease-related information with hyperedge-aware hypergraph convolutional networks. We introduce two loss functions: an atlas-based contrastive loss and a class-consistency loss to guide the disentanglement processes. We evaluate our model on the public Autism Brain Imaging Data Exchange (ABIDE) dataset to demonstrate the effectiveness of the proposed model and investigate the optimal combination of brain atlases. Our results shed new light on the importance of exploiting the relationship among by disentanglement for improving multi-atlas disease diagnosis. In addition, our model provides deeper insights into disease interpretability, including atlas properties and critical brain regions. Our code is publicly available at <a href="https://github.com/MortonHao/BrainUMA">https://github.com/MortonHao/BrainUMA</a>.</p> Graphical abstract <p></p>