Understanding disease similarity is pivotal for unraveling pathological mechanisms and facilitating drug repositioning. However, conventional computational approaches often fall short in effectively capturing the complex multi-molecular interactions and subtle surveillance signals present in biological systems. To address this limitation, we propose RGMI, a graph neural network framework based on dynamic weight calculation and alternating learning strategy, which aims to accurately predict potential disease associations by integrating gene network, miRNA network and disease data. RGMI employs graph convolutional network to encode features derived from gene network. Additionally, it incorporates an advanced dynamic attention mechanism to model the heterogeneous dependencies among miRNAs. By comparing losses across multiple views, RGMI enhances the consistency of gene-miRNA associations, thereby producing high-quality representations. To construct robust disease representations, RGMI introduces a weighting aggregator that dynamically selects biologically relevant features. It further captures intricate interactions between disease features using bilinear transformations, which are then utilized to generate prediction scores. We have developed a regularization scheme that mitigates modal competition, improves the training process, and enhances both training stability and generalization performance through periodic global multimodal integration learning. Experimental evaluations demonstrate that RGMI significantly outperforms existing state-of-the-art methods across multiple key metrics. These results highlight its superior capability in predicting disease similarity.