Graph Mask Contrastive Autoencoder for MiRNA-Disease Association Prediction

MiRNA-disease association (MDA) prediction plays a crucial role in advancing our understanding of disease mechanisms and developing new therapeutic strategies. However, existing methods face two major challenges: high dependency on the reliability of input data, and difficulty in effectively extracting complex MDA patterns from multi-source graph data containing noise and incompleteness. To address this, we propose a novel masked graph autoencoder contrastive learning framework, GMCAEMDA, which leverages self-supervised learning to capture features from noisy and incomplete graphs. First, we introduce a graph mask autoencoder to reconstruct node features, employing a masking encoding and re-masking decoding strategy to enhance the model’s ability to learn effective node representations from noisy and incomplete data. Then, to further improve the model’s capacity to learn unmasked node representations, increase robustness, and reduce overfitting during training, we construct a contrastive feature matching module based on a teacher-student paradigm. Experimental results on two versions of the HMDD databases demonstrate that GMCAEMDA outperforms existing methods, achieving significant improvements in miRNA-disease association prediction.