To characterize atypical brain dynamics under diseases, prevalent studies investigate functional magnetic resonance imaging (fMRI). However, most of the existing analyses compress rich spatial-temporal information as the brain functional networks (BFNs) and directly investigate the whole-brain network without neurological priors about functional subnetworks. We thus propose a novel graph learning framework to mine fMRI signals with topological priors from brain parcellation for disease diagnosis. Specifically, we 1) detect diagnosis-related temporal features using a "Transformer" for a higher-level BFN construction, and process it with a following graph convolutional network, and 2) apply an attention-based multiple instance learning strategy to emphasize the disease-affected subnetworks to further enhance the diagnosis performance and interpretability. Experiments demonstrate higher effectiveness of our method than compared methods in the diagnosis of early mild cognitive impairment. More importantly, our method is capable of localizing crucial brain subnetworks during the diagnosis, providing insights into the pathogenic source of mild cognitive impairment.