Prediction of lncRNA-disease association based on heterogeneous graph contrastive learning

Long non-coding RNAs (lncRNAs) play pivotal roles in human physiology and disease pathogenesis; however, accurately predicting their associations with diseases remains a significant challenge due to the limited predictive accuracy of existing methodologies. To address this issue, we introduce HGCL-LDA, an innovative framework that utilizes self-supervised graph contrastive learning to achieve high-precision predictions. Our methodology consists of three key steps: First, we construct a heterogeneous graph (HLD) by integrating diverse similarity metrics and association data. Second, we generate multi-view graphs through advanced data augmentation and perturbation techniques, followed by the extraction of node embeddings using a GCN encoder. Finally, we perform contrastive learning by constructing positive and negative sample pairs and employ XGBoost to predict potential lncRNA-disease associations. Extensive experiments on four datasets demonstrate HGCL-LDA’s superior performance over state-of-the-art models.The model was applied to lung, gastric, and liver cancers, predicting the top 15 lncRNAs potentially associated with each cancer type. Among these predictions, 14 lncRNAs for lung cancer, 13 for gastric cancer, and 12 for liver cancer were experimentally validated, respectively, highlighting its biomedical relevance. These results not only demonstrate the robustness and reliability of our framework but also highlight its potential to uncover novel lncRNA-disease associations, thereby advancing our understanding of disease mechanisms and contributing to the development of targeted therapeutic strategies.