Metabolic diseases, such as diabetes, pose significant risks to human health due to their complex pathogenic mechanisms, complicating the use of combination drug therapies. Natural medicines, which contain multiple bioactive components and exhibit fewer side effects, offer promising therapeutic potential. Metabolite imbalances are often closely associated with the pathogenesis of metabolic diseases. Therefore, metabolite detection not only aids in disease diagnosis but also provides insights into how natural medicines regulate metabolism, thereby supporting the development of preventive and therapeutic strategies. Deep learning has shown remarkable efficacy and precision across multiple domains, particularly in drug discovery applications. Building on this, We developed an innovative framework combining graph autoencoders (GAEs) with non-negative matrix factorization (NMF) to investigate metabolic disease pathogenesis via metabolite-disease association analysis. First, we applied NMF to extract discriminative features from established metabolite-disease associations. These features were subsequently integrated with known relationships and processed through a GAE to identify potential disease mechanisms. Comprehensive evaluations demonstrate our method’s superior performance, while case studies validate its capability to reveal pathological mechanisms in metabolic disorders including diabetes. This approach may facilitate the development of natural medicine-based interventions. Our data and code are available at: https://github.com/Lqingquan/natural-medicine-discovery.