Objective Gut microbiota dysbiosis has been implicated in various diseases, making the identification of microbe–disease associations essential for understanding disease mechanisms and discovering disease-associated microbes. Computational approaches have been widely used to predict microbe-disease associations and prioritize promising candidates for experimental validation. However, existing methods often suffer from insufficient utilization of heterogeneous biological information, particularly the complementary structural and semantic information derived from association networks and similarity data. Moreover, the sparsity, noise, and incompleteness of known associations further limit prediction performance and generalizability.Methods that primarily rely on a single information source or simple feature fusion may fail to capture the complex relationships between microbes and diseases. In this paper, we propose a graph-enhanced contrastive learning framework for microbe–disease association prediction, GECLMDA, which integrates complementary structural and semantic information to obtain informative node representations and improve the prediction of potential associations.Methods Our method derives informative node representations by integrating structural and semantic learning. Specifically, graph convolutional networks are employed to capture structural relationships between microbes and diseases, while a deep sparse autoencoder is used to extract latent semantic features from similarity information. The structural learning component captures relational information from graph-based representations, whereas the semantic learning component extracts complementary information embedded in microbial and disease similarities. Unlike methods that directly contrast multiple graph-derived representations, GECLMDA integrates graph-based relational information before performing cross-representation alignment. Contrastive learning is then introduced to align the structural and semantic views, allowing the model to jointly exploit relational structures and latent similarity patterns. Finally, the resulting embeddings are fused and used to predict potential microbe–disease associations. The method is evaluated on three benchmark datasets, HMDAD, Disbiome, and Peryton, using five-fold cross-validation under different evaluation settings. Ablation experiments and case studies are further conducted to assess the contributions of key model components and the ability of GECLMDA to identify potential disease-associated microbes. In addition, ablation experiments are performed to investigate the contributions of the major model components, and case studies are conducted to examine the ability to identify potential disease-associated microbes.Results Our method achieves competitive performance against most baseline methods across the evaluated settings. Under global five-fold cross-validation, GECLMDA obtains AUCs of 0.980 3±0.006 9, 0.944 6±0.008 3, and 0.954 6±0.006 7 on HMDAD, Disbiome, and Peryton, respectively. Ablation experiments indicate that structural representation learning, semantic representation learning, and contrastive alignment provide complementary contributions to prediction performance. Case studies further suggest that our method can serve as a potential tool for predicting disease-associated microbes.Conclusion The proposed framework effectively integrates structural and semantic information through graph-enhanced contrastive learning and achieves competitive prediction performance across multiple benchmark datasets. GECLMDA provides a potential computational tool for microbe–disease association prediction and the identification of disease-associated microbes, offering useful candidates for further biological investigation.