In brain disease diagnosis, multimodal medical data often faces challenges such as inter-modal data drift, inconsistent feature distributions, and insufficient prediction robustness. To address these limitations, we introduce MANGL, a graph learning framework designed to integrate multimodal feature alignment and masked random noise perturbation. We enable multimodal feature alignment through the framework’s cross-modal semantic mapping network, thereby effectively resolving inter-modal drift and inconsistent feature distributions. Concurrently, we integrate a masked random noise perturbation strategy within the graph convolution modules to enhance disease prediction robustness. We conduct comprehensive experiments on three multimodal datasets encompassing Alzheimer’s disease and autism spectrum disorder, and the results show that MANGL outperforms state-of-the-art baselines in both accuracy and AUC. These results confirm the framework’s efficacy in resolving inter-modal data drift, enhancing model robustness, and enabling precise diagnoses of complex neurological disorders.