Metabolites are small molecules produced during organism metabolism, with their abnormal concentrations closely linked to the onset and progression of various diseases. Accurate prediction of metabolite-disease associations is crucial for early diagnosis, mechanistic exploration, and treatment optimization. However, existing algorithms often overlook the integration of node features and neglect the impact of different hop domains on nodes in the processing of heterogeneous graphs. Furthermore, current methods solely rely on random sampling for selecting negative samples without considering their reliability, thereby compromising model stability. A novel metabolite-disease association prediction model, GMAMDA, is proposed to address these challenges. GMAMDA integrates adaptive hardness negative sampling, adaptive graph multiple convolution techniques, and a multiheterogeneous graph fusion strategy to forecast potential metabolite-disease associations. Initially, by computing multisource similarity information for metabolites and diseases, multiple heterogeneous graph networks are established for metabolite-disease association networks. Subsequently, the adaptive graph's multiconvolution mechanism is employed to generate feature-rich node representations across various heterogeneous graphs by dynamically leveraging information from different hop neighborhoods. The model then utilizes an adaptive hardness negative sampling approach based on principal component analysis to select negative samples with the highest information content for training, enabling the prediction of potential associations between new metabolites and diseases. Experimental findings demonstrate that GMAMDA outperforms state-of-the-art methods across various evaluation metrics, including AUC (0.9962 ± 0.0014), AUPR (0.9967 ± 0.0009), and accuracy (0.9733 ± 0.0042). Case studies focusing on Alzheimer's disease and kidney disease further validate GMAMDA's clinical potential in predicting metabolite markers.